CCTV News:The Central Meteorological Observatory issued a cold wave blue warning at 10: 00 on December 9:

　　Affected by the cold wave, it is estimated that from 14: 00 on December 9 to 08: 00 on December 12, the temperature will drop by 6 ~ 10℃ in parts of the southeastern part of northwest China, northeast China, North China, Huanghuai, Jianghan, Jianghuai, Jiangnan and northern South China, and the temperature drop in eastern Heilongjiang, southern North China and Huanghuai can reach more than 12℃. On the morning of the 12th, the lowest temperature line of 0℃ will be pushed southward to the area from central Hubei to northern Jiangsu and Anhui.

　　It is estimated that from 14: 00 on December 9 to 14: 00 on December 11, the temperature in most parts of Northeast China, southern North China, most of Huanghuai, most of Jianghan and western Jiangnan will drop by 6 ~ 10℃ successively, among which the temperature in some areas of eastern Heilongjiang, southwestern Shanxi, western Huanghuai and northern Hubei can reach above 10℃.

　　Defense guide:

　　1, the government and relevant departments in accordance with their duties to prepare for the cold wave;

　　2. Pay attention to adding clothes to keep warm;

　　3. Take certain protective measures for tropical crops and aquatic products;

　　4. Prepare for the wind.

　　Bath center, 24-hour ATM bank, underground parking, Internet cafe, KTV… … In recent years, some tourists choose these cheap and convenient places to spend the night in order to save money when traveling abroad. Recently, however, Red Star journalists found that some netizens shared their "coup" of saving money by staying overnight on a social platform, and even chose the overnight place in the emergency department of the hospital. They also said that big hospitals such as the top three hospitals are five-star hotels with intimate service and comfortable recliners.

　　In this regard, some medical staff working in the emergency department told the Red Star journalist that in the past work, people who did come to see a doctor by pretending to be sick generally did not deliberately expose them, and they would be treated according to normal diagnosis and treatment. The inspection should be advised to go back to rest and persuade them to leave. It is considered to be very unqualified for this kind of behavior that takes up the queue for medical treatment and medical resources. "If there are many such behaviors, the hospital will issue relevant regulations in the future, which will add trouble to those who really come to see a doctor."

　　▲ Screenshot of the post of saving money by going to the emergency department of the hospital for the night.

　　Bloggers send "coup" to save money by traveling;

　　"You can pretend to be sick and go to the hospital for emergency treatment for the night."

　　Netizens criticized "guiding others to occupy medical resources"

　　Recently, a blogger posted on a social platform that if you can’t book a room for travel, you can register to stay in the emergency room of the hospital. According to the screenshot of the net transmission, the blogger created his own "small coup" to save money for overnight travel, and emphasized that big hospitals such as the top three hospitals are five-star hotels with intimate service and comfortable recliners.

　　The blogger also described in detail a series of procedures for registering and seeing a doctor in his post: registering at the emergency window of the hospital after 10 pm, taking blood pressure and temperature at the triage desk, washing after "confirming that he is not sick", lying that he coughed and was out of breath when seeing a doctor, and asking the doctor to prescribe oxygen instead of medicine, and staying in the hospital for the night at the price of 3 yuan per hour. The time of oxygen inhalation can be determined according to his sleep habits. "A total of 30 yuan can spend the night on the hospital lounge chair, and the nurse will help look after the luggage."

　　On October 9, when Red Star journalists searched the above-mentioned bloggers’ accounts on the platform, they found that they could not find the relevant accounts. However, by searching keywords on the platform, the reporter found that other bloggers still posted similar posts. According to a blogger, his company organized a very remote (local) physical examination, and decided to stay in the hospital because he couldn’t go back that day. At that time, the doctor immediately opened a hospital for him, or a single room.

　　▲ Another netizen shared the post "Where can I spend the night without staying in a hotel?"

　　The blogger also said in his post that the hospital fee is basically about one day in 60 yuan, but only when the beds are free, and a sick note can be issued. Later, when the blogger went to Hangzhou to play, he had a physical examination in a local hospital because he couldn’t find accommodation. "I was given an accommodation. I was called up at 6: 30 the next day and took a color ultrasound. I let me go. I spent 220 yuan before and after. The accommodation price in Hangzhou was not lower than that in 600 yuan that day."

　　Since the post of "saving money by spending the night in the hospital emergency room" was exposed, it immediately triggered a hot discussion among netizens. Many netizens left a message in the comment area, "This is guiding others to occupy medical resources, which will seriously interfere with the normal order of hospitals and the work of medical staff"; "Emergency overnight not only wastes medical resources, but also affects patients who really need emergency services. Can you sit still in the rest area here? " Some people also said, "Not to mention the need for doctors to judge the condition, all kinds of patients will be received all night, as well as the sounds of various instruments, all kinds of painful screams, and the lights will shine like daylight, so you won’t go even if you are asked to go" … …

　　▲ Reply from the platform involved

　　On October 10th, the Red Star journalist contacted the customer service (professional assistant) of the platform involved. The other party said that in this case, you can click the reporting option on the platform, and provide the relevant materials of the whistleblower according to the corresponding prompts on the page, and the corresponding Commissioner will follow up the verification. "The platform will take one or more of the following measures based on the facts independently identified by it, such as: limiting the scope of display, banning accounts, banning accounts, and pursuing legal responsibilities according to law."

　　Emergency medical staff:

　　Encountered pretending to be sick to see a doctor.

　　But it is almost impossible to achieve the goal of staying overnight at a low price.

　　"Emergency departments at home and abroad are roughly divided into five major areas, including pre-hospital emergency triage area, hospitalization observation area, rescue area, emergency inpatient ward (including general ward and emergency intensive care unit) and infusion area." Xiao Wei, a medical staff member in the emergency department of a third-class first-class hospital in Guangzhou, told Red Star that the day shift and night shift in the domestic emergency department are similar. The day shift is the normal administrative time, around 8: 00 -18: 00, and the rest of the time is night shift. "After the outpatient clinic is closed, the emergency department becomes busy, usually at noon and evening, as well as on the eve of holidays and during holidays."

　　Regarding the bed problem in the emergency department, Xiao Wei said that there will be beds in the rescue area and the observation area, but the number is not very crowded, and they all need to go through the formalities. "The emergency inpatient ward, like the inpatient department, needs hospitalization procedures to live. There are 5 beds in our infusion area, which are used by patients who need infusion. After transfusion, they must get up and lie down for patients who need it later. People who don’t need infusion are not allowed to lie down. " In addition, Wei also said that the general process of treating patients in the emergency department is: when patients arrive in the emergency department, there is a triage desk, and the nurses at the triage desk will evaluate the patient’s condition, roughly grade the condition by asking questions, measuring heart rate and blood pressure, and queue up for treatment according to the grading system.

　　"Patients who pretend to be sick have also encountered some. The triage nurse generally does not refuse to consult, and the registration will be hanged. However, if it is recognized that he is pretending, he will be told that he must queue up according to the order of diagnosis and treatment and cannot jump the queue. After it’s his turn to see a doctor, the doctor will judge his condition. It’s really fake, and he won’t open an oxygen inhalation or a bed for you to lie down for one night. " Wei said, because doctors also know that he is responsible for the patients in every bed he arranges, he also needs to make records and shift the next doctor, and so does the nurse. "It is easier for people who pretend to be sick to see it. Ask more questions, do some relevant physical examinations, open tests for him, and he will refuse to check. However, (medical staff) generally will not deliberately expose the illness, so they will still be treated according to normal diagnosis and treatment. The inspection should be advised to go back to rest and persuade them to leave. Even if there is a free bed, it is impossible to give him a night. "

　　In addition, regarding the aforementioned problems of being drunk and being hospitalized for physical examination, Xiao Wei admits that the control of being hospitalized in the emergency department is stricter than that in the general outpatient department, so the situation of being hospitalized in the emergency department through being drunk or being hospitalized for physical examination hardly exists in the emergency department. "Unless the condition of drunkenness is very critical, it will definitely be hospitalized for treatment. “

　　"This online ‘ Planting grass ’ I was a little shocked by the situation. It is not excluded that some people have done this and succeeded, but this situation generally does not appear in busy public hospitals. However, many hospitals are expanding now, and the new campus of some hospitals is indeed well decorated and equipped, and it is not excluded to open beds to live in order to generate income; Moreover, some young doctors are inexperienced. If people who come to see a doctor strongly demand to stay overnight even if they have not detected any problems, there are also empty beds. For safety reasons, doctors may think that he may have mental illness and let him stay for one night. " Wei told reporters that they had had patients with mental and sleep problems before, and only asked for oxygen and overnight stay, but the patient did have mental problems. "In short, the hospital emergency department generally does not refuse to consult, but it must abide by the order of diagnosis and treatment, because emergency patients will definitely give him priority in cutting in line. Moreover, sometimes in an emergency, such as a new patient who needs to be rescued, if there is a lack of beds, the emergency department does not rule out letting patients with mild illness give up their beds, take wheelchairs or seats first, and wait until the beds are turned around before returning them to patients with mild illness. "

　　"Patients with colds and fever usually need to leave after the fluid is lost, and they are not allowed to stay in the emergency room. If medical staff find that someone is pretending to see a doctor, they will continue to urge them to leave. " The medical staff in the emergency department of a second-class hospital in China also told reporters that in the past, there were cases in which patients broke into the emergency room for the night by themselves in the middle of the night, and the medical staff would drive away the people who occupied the beds during the morning rounds. "Under normal circumstances, all patients who saw a doctor in the emergency department, as long as they did not apply for hospitalization, would ask the patients to leave after the treatment, and would not be allowed to stay overnight. It is almost impossible for people without symptoms to want a bed in the emergency department through the normal medical treatment process to achieve the purpose of staying overnight at a low price in the hospital. "

　　The medical staff also said that doctors will diagnose the specific symptoms of patients and will not blindly meet the requirements of patients. I just want to take oxygen when I cough, so don’t prescribe medicine ’ In fact, not everyone is suitable for oxygen inhalation when coughing. Even if oxygen inhalation is allowed, it is only allowed for twenty or thirty minutes. After simple treatment, the patient needs to leave, and it is impossible to stay overnight. "

　　Red Star Journalist Luo Mengjie Intern Wang Zixuan

Introduction:

On August 24th, the latest issue of Horizon’s "Hello, Developer" Autopilot Technology Special Session was successfully broadcast in the Zhidong Open Class.

The special session was given by Dr. algorithm engineer Zhu Hongmei, Horizon Perception, with the theme of BEV Perception Scheme and Deployment Practice Based on Journey 5. This article is a record of the lecture session of this special session. If you need live playback and Q&A, you can click

Good evening, friends of the open class of Zhidong. I’m Perception algorithm engineer Zhu Hongmei from the horizon, and I’m honored to share with you today "Ji"BEV awareness scheme and deployment practice in Journey 5》。 I will introduce it from five aspects.

01

General introduction of BEV perception framework

With the release of Journey 5 in 2021, we designed the sensing prototype scheme of BEV based on Journey 5. The input is vehicle-mounted multi-angle image sequence or multi-mode signal, and the fusion of time and space dimensions is carried out within the network, so that the neural network can output the dynamic and static sensing and prediction results from a bird’s-eye view.

Below is a frame diagram.

The green background part represents the signal processing of pure vision, and the input is the images of each camera. After the first stage of model extraction, the image features are transformed into multi-view spatial fusion from BEV perspective. Then, the features of BEV are fused in time sequence, and then sent to the second stage for feature extraction on BEV; Finally, input it to the head part for the output of perceptual elements. If there is a Lidar in the system, it will also be connected to the Lidar point cloud. In the early stage, we will do rasterization coding processing, and we will also do time sequence fusion. Finally, we will do intermediate fusion with visual BEV features, and also send them to the second stage to output the perceptual elements on BEV.

The blue part is the closed-loop iteration that has participated in many rounds. Students who pay attention to the horizon should know that at the Shanghai Auto Show in April this year, this part of the content has supported the closed-loop demo display of real cars. In recent six months, we have expanded the function of "dynamic obstacle detection" in sensing elements. Based on it, we connected the head of Transformer to do the Track id, speed estimation and trajectory prediction of dynamic obstacles. The dotted line part is currently under development, and we also want to make full use of the dynamic and static sensing results on BEV to output the interactive elements of dynamic and static sensing, such as behavior prediction.

There are two main modes of pure vision scheme here. One is the 11v integration scheme for double Journey 5; The other is a single journey chip-oriented scheme, in which 7v for driving and 5v for parking are integrated. Today’s sharing mainly focuses on the integration of time and space in the visual scheme, the analysis of perceptual elements, and the introduction of dynamic, especially dynamic perceptual prediction end-to-end scheme and deployment.

02

Spatio-temporal fusion module and chip deployment

Firstly, it is introduced that the four-plane space fusion method is adopted at the end of the plate at present. It takes the xy plane of the self-propelled coordinate system as the basic plane, and adds three height planes on this basis, hoping to obtain more scene image information with different heights. Time series fusion is relatively simple. We will cache the features after time series fusion at the last moment, align the historical frames to the current frame by using the self-driving motion between frames, and fuse them with the spatial fusion features of the current frame, and then deconvolution them and send them to the following network structure for perceptual output.

So why do we use this scheme?

In the early stage of research and development, we also compared different spatial fusion methods. In the early stage, in order to quickly get through the process, we adopted the simplest method based on single plane IPM to get through the perception of the whole framework. After that, I went to find out the depth-based fusion and transformer-based fusion scheme on Journey 5. The assessment dimension is the delay in Journey 5, assuming that the prerequisite is to be able to independently run this model above 30fps on the single core of Journey 5 to design the corresponding network parameters.

There are two key parameters that affect the perceptual performance of BEV, one is the resolution of image feature in one stage, and the other is the fusion resolution on BEV. The fusion resolution on BEV is relatively direct, and the higher the resolution, the better the perception of some small targets and point elements. Image resolution determines the amount of information input in two stages.

Here is a lane line simulation diagram, and the physical dimensions of each lane line are the same. Because it is perspective projection, the farther the lane line is in the image, the fewer pixels it will have. BEV indexing image features is generally a process of reverse realization, while BEV is generally designed with equal resolution, such as 0.6 and 0.4 m Grid resolution. When the distant objects in the scene are indexed by the coordinates of BEV, the physical resolution of the pixels in the image is already lower than that on BEV, and at this time, feature duplication will appear on BEV, which is a common stretching phenomenon. Reflected in the perception, for example, the detection box may have the phenomenon that the target splits into a string, so the feature resolution in one stage is also very important.

We grab the middle column of pixels to simulate the contrast of Z-value changes between two adjacent rows of pixels in 2M image and 8M image at 50m. It can be seen that the physical range of the low-resolution image corresponding to two adjacent rows of pixels has changed by more than 2 meters. For a high-resolution image, its adjacent two pixels are only about 0.6 meters. Therefore, when looking at the fusion scheme, we should not only consider whether a single task meets certain accuracy requirements, but also consider the final deployment, because it is multi-task and different perceptual elements output, so we hope that the resolution of one-stage feature and BEV fusion can be higher.

From this point of view, the current structure based on Transformer can meet the requirements of Journey 5, single core and 30+FPS, and the first-stage resolution and the second-stage fusion resolution are relatively low, and the delay is too high under this condition. However, the depth-based fusion method is lower than single plane in one stage. In addition, one reason for not adopting it is that the depth-based deployment has made some specific improvements for Journey 5, reaching a two-phase delay of 9.6 ms. In the stage of prototype verification, the flexibility of carrying different numbers of cameras on the car is not enough, so we adopt the plane mapping fusion method at this stage. In the early stage, we used a single plane, but when we actually did the closed-loop verification, we found that the generality and robustness of camera parameters were not enough, and it was sensitive. Therefore, we improved it in the further research and development process. Through various tests and verifications, we finally adopted the four-plane fusion method, which can achieve a good balance between performance and delay.

For time series fusion, different time series fusion methods will also be compared in the early stage of research and development. The basic method is to align the historical BEV features to the current frame and use different fusion methods to do time series fusion. Here is also a comparison of the fusion of different frames, with the experimental data accumulated at that time to do 10 frames of verification. In the case of 10 frames, there is an improvement of 3.5 points compared with only space. For latency, it is only increased by 1ms because it is deployed in recurrent mode.

Similarly, we also compare the two methods of obtaining historical frames. One is to intercept historical information at equal time intervals, and the other is at equal space intervals. Equal time interval is to use the system timestamp to index and read the corresponding frame. Equal space interval is to calculate the translation distance with the movement of the vehicle, and then get the frame. Finally, we found that the performance of equal space interval of 6 frames is similar to that of equal time interval of 10 frames, and they can also be combined, but the improvement of more fused frames is not great. Therefore, in order to simplify the deployment, we currently adopt the time series fusion method with equal time intervals.

There may be some differences in the training and deployment of time series fusion. One is the Odometry used when aligning two adjacent frames. In order to make the training model performance consistent with the real vehicle deployment, one principle to be observed is that the model input should be consistent with the board signal as much as possible. We use Odometry of wheel speedometer to do model training, and do tests on Odometry obtained by different methods. It is found that if it is consistent with the inter-frame interval during training, then this factor has little influence. In addition, for example, historical frames are obtained at 100ms intervals, and the model trained with 10fps training data is tested at 10fps, 15fps and 30fps. The conclusion is that if there is a big difference between deployment and training, the index gap will be even bigger. The actual system frame rate is affected by the real-time load of the system, and it may not be completely aligned no matter how to set the inter-frame interval during training. Generally, the perimeter perception can be around 15fps, which can basically meet the functional requirements, and the range of 1.5 points is also acceptable. This part can also make the distribution of time series training data more abundant through some data enhancement methods, and alleviate such problems.

The other is that when the time series is being trained, if there is no specific optimization, with the increase of the number of frames, some OOM(Out of Memorry) problems may occur in the time series data, which leads to the slow training speed. The training method is to adopt the feature cache strategy to alleviate or avoid such problems. However, when deploying on the board side, considering the memory limitation of the board side and the bandwidth loaded back, the most streamlined strategy recurrent mode is finally adopted on the board.

03

Analysis of static and Occupancy perception elements

An important component of static perception is truth value. How did the truth value on BEV come from?

From the very beginning of research and development to the present, three static truth values have been used. At first, high-precision maps were used, and now the truth production link based on multi-mode is used on a large scale. At present, the local mapping results of pure vision are also involved in the training. The quality of the drawings is developed by our company’s special 4D-Label team. At the special session of horizon "Hello, developer" autopilot technology, Dr. Sui Wei, the head of horizon 4D labeling technology, also shared the "4D labeling scheme for BEV perception". If you are interested, you can review it.

For the algorithm, the algorithm students should be more involved in the formulation of labeling rules. Even if the truth value is brushed by the big model, the big model needs the truth value, so the labeling rules need the deep participation of the students of the perceptual algorithm, especially the actual test scenario is very complicated. A reference condition of labeling rules is the Lidar point cloud base map obtained by mapping, which is an intensity map on BEV, and then it can be labeled by referring to images or videos. For example, lane lines should be marked with colors, some functional attributes and some downstream functions that are necessary and cannot be obtained from the Lidar bottom map, all of which need to be combined with images. For pavement signs, they are usually unmarked rotating frames, and for those that need a clear direction, they need to be marked in an orderly sequence. The elements of parking are basically to mark parking spaces, wheel gears, speed bumps and so on by marking key points.

The output of the model corresponds to the annotation. The marked information will be preprocessed to generate the true value of the model. Only when the truth value is supervised output can the downstream functional requirements be met. The lane line output is the position and attribute of the example point column; The intersection point is the position and type of the point; Curbside is also the instance position and type; Pavement discrete signs are detection frames and types. These static perception elements are very rich. For the model scheme, there is a classification learning of foreground extraction, a regression learning of accurate position, and a learning of example features of lane lines. On the whole, the static perception scheme is relatively conventional. When analyzing, line elements such as lane lines will extract the foreground through the score of Class Map, then cluster the points through the Embedding similarity, and reverse the attribute features to index the corresponding features of these points.

For lane lines, in the actual software analysis, we should consider the delay on the software, roughly estimate the direction of lane lines through vertical points according to the preset, and do one-dimensional search along this direction to shorten the post-processing time of lane lines. Now our scheme is newly optimized, and it takes less than 3ms to deal with this part. The analysis of pavement signs, detection points or boxes at intersections is a set of CenterNet: card score→maxpolling→TopK→ attribute extraction → output.

For parking elements, parking spaces will output orderly corner position, occupancy and parking space type direction; The wheel stopper is an orderly end position and type; The underground locks and columns under the parking lot are all output by detection frames, which are the same as the analytical ideas of the previous points and line frames. The parking space here has two small head, one of which is to obtain the overall attributes of the parking space, such as its occupancy, and whether the parking space type is parallel or vertical. There is also a Local head to accurately estimate the corner of the location, because the accuracy of this corner of the location is very high. After getting a more accurate one, calculate the point matching with the rough corner position obtained under global. Because this point is orderly, you can just card a distance threshold. After matching, you can get an accurate corner coordinate by replacing the corner coordinate.

This is our performance. The 2D perception scheme of horizon has been polished for a long time before, which is equivalent to having a good reference baseline for BEV perception.

Compared with the 2D scheme, lane lines and road signs have a great improvement in ranging, such as lane lines in different distance sections, arrows and stop lines of road signs. For parking spaces, Horizon also has a pre-fusion scheme. Before being sent to the network, four fish-eye images are fused by IPM to get an IPM image covering 360 degrees around the car, and then integrated into the network to output parking spaces. Now BEV is introduced in the previous framework, and feature are fused inside the model. At present, there are also great improvements in these three class libraries.

Take a look at this video, which is a visual display of a high-speed lane line, with a wide dotted line, and then get off the ramp. Here is a diversion point, a change point between reality and reality, and the arrow represents the direction.

For the perception of parking space, a parking scene is chosen here, where the red line represents the entrance line, the yellow X represents the occupation, the yellow box is the pillar of the parking lot, and the grass green is the wheel stopper.

Let’s talk about the perception of Occupancy. First of all, Occupancy is a 2D occupancy graph based on Journey 5, which is a binary graph, where 0 means unoccupied and 1 means occupied. We call it Freespace internally. The traditional Freespace is image segmentation and post-processing to calculate the drivable area under BEV. It is more natural to do it on BEV, and output such a binary image directly on BEV. Its true value depends on Lidar segmentation model+detection model+threshold of cloud card height of multi-frame aggregation point to obtain some obstacles on the ground. Lidarseg mainly obtains the basic surface first, and takes the road plane as the drivable area; Lidar3D detection is to supplement the true value of dynamic obstacles in the white list.

The other output is Vismask. It mainly describes the area that the vehicle can see in the image at the current moment, and also uses Lidarseg model and 3D detection model. Image segmentation combined with back projection of point cloud can get the points that can be seen under BEV, calculate the ray according to the semantics and camera parameters, and then calculate the surface paved by the points that can be seen in the image in the self-propelled coordinate system, and then flatten them in the height direction of point cloud to get the visible Grid points on BEV. Why do you need 3D detection? Because when the ray hits the point cloud, it is a surface. We want the whole object of the surface boundary to be visible, so we need a 3D detection frame to supplement the object of the boundary, thus obtaining the true values of these two tasks. Perception model is a relatively simple binary classification model. Vismask’s 0 means invisible and 1 means visible.

Because these two tasks are dense outputs, when actually doing this on the board, its bandwidth occupation is relatively large. Usually, we will use a Run Length Encoding (RLE) to compress the result. For example, Freespace, as shown in the following figure, has the same values in a connected region. Run-length coding is used to reduce the bandwidth of the perceptual results transmitted from BPU of Journey 5 to CPU, and a final output binary image can be obtained by doing a back analysis in CPU.

Here is a visual video, and the bar on the left is a response to the fence. This is a specially built evaluation set, including strollers, fallen cones and so on. Here, Freespace is mainly to supplement the perception of non-standard obstacles. In addition, it is also useful to feed back the representation of roadside by Freespace when the downstream closed loop is used.

04

Dynamic Perception Predicting End-to-End and Chip Deployment

Let’s introduce the end-to-end scheme of dynamic sensing prediction and chip deployment.

The end-to-end model, as shown in the previous frame, is followed BEV3D several layers of Transformer decoder to output dynamic end-to-end elements. The multi-scale features extracted by BEV in two stages are reused, and there is a detection head of BEV3D, which regresses to the heatmap and the offset to the precise position to get the possible target position.

For Transformer head, it depends on how the Q/K/V of its input comes from the first layer. There is a construction part of the input. First of all, "q" is divided into detection query and Track query. Detect the position embedding； of query, which is the position parsed from BEV3D head and obtained by MLP; Context embedding is learnable and can be initialized randomly.

For the position embedding of Track query, the position of the previous frame will be used, and the self-vehicle motion between frames will be aligned to the current frame to eliminate the self-vehicle motion. Then, the position of the current frame is estimated by combining the speed estimation result of the previous frame itself, and the position embedding of Track query is obtained by MLP. Its context embedding was also randomly initialized and learnable at the beginning, which resulted in "Q".

"K" and "V" For BEV, there is a fixed mapping relationship between the pixel coordinates on BEV and the coordinate positions under VCS. Therefore, according to the position of the reference point, its pixel coordinates on multiple dimensions of BEV can be calculated reversely. Query also sorts from 0 to N-1, and it has an index, so that the difference between the two coordinates can be calculated into the mapping relationship between them. Then the warping operator, that is, the grid_sample operator, is used to reversely index the BEV region features, and then "K" and "V" are generated through MLP, so that the end-to-end head input is constructed.

One of the differences in Transformer here is that it improves the quantization and efficiency on Journey 5, and removes the MatMul matrix multiplication operator. Here, the commonly used Deformable attention is replaced by warping.

The output layer of end-to-end head is actually several MLPs. Among them, the output of trajectory and speed will pass through Memory bank in order to obtain historical information. The point of the trajectory output here is also the displacement between frames, and the velocity describes the position change of unit time between frames in the future. We think that these changes are not abrupt, but smooth, so we use historical information.

The embedding of history is composed of two parts. The first part is the Oddity of N historical frames and the embedding of query output by Transformer decoder of N historical frames will be stored in Memory bank. Query of the current frame, and pay attention to the time sequence of the current frame and the historical frame for each target. After that, self-attention will also be done among the targets, and finally we define it as Temporal embedding. In addition, speed embedding is added. How did it come from? That is, the position of historical frame N is transferred to the current frame through the odometry of the vehicle motion, and the position of the historical frame under the current frame is obtained, so as to eliminate the vehicle motion and obtain the inter-frame offset. Combining with the frame rate of the system, a time difference can be derived, and the historical inter-frame speed can be estimated. Similarly, after passing an MLP, the embedding of the speed can be obtained. By making a simple concat, the final historical embedding can be obtained and input to the trajectory.

After the model is output, some of it will be analyzed in CPU. For Track query and Detect query, the model output itself will analyze the position, length, width, height, orientation angle, Score confidence and so on. The function of track management is to screen out the Track query of the next frame. Two thresholds are set here. A higher threshold is to filter out the tracked query with higher confidence from all the queries, that is, the tracked query, and identify the newly generated target in the current frame. For the newly generated target, a current maximum id value of +1 is given, and its disappearance time is initialized to 0. For setting a low threshold, I want to recall those targets that have been tracked before, but have low confidence in the current frame due to occlusion and other reasons. If it can be greater than this threshold, and the disappearance time is not greater than the set time, it will continue to be preserved. If the time of disappearance (disp_times) has been greater than the threshold of the number of frames that can be dropped, we will definitely drop this target in this frame.

For the analysis of velocity, the model itself can be directly output, and kinematics post-processing will be added after it comes out. The "z" here is the length, width, height and speed (attribute) of the previous output, which we call observation. The model will also learn from the gain matrix of Kalman, avoiding calculating "k" on CPU, and finally giving the filtered state quantity to the downstream. Compared with the unfiltered operation, this operation has a 19% improvement in speed accuracy. From the visual performance, the speed does appear smoother after this. When updating the embedding of Track query in the next frame, the historical embedding generated in Memory bank will be used. The details of QIM will not be expanded, so you can refer to the implementation of MOTR.

Take a closer look at the performance of the end-to-end model. End-to-end comparison is only detection, tracking and speed estimation through post-processing. The baseline detection compared here is already 3D detection on BEV. Compared with the original 2D detection, it has obvious advantages in ranging. Blue and green are the vertical error percentages of 2D detection in different distance segments, and red and purple are BEV3D. It can be seen that the end-to-end model is more advantageous than BEV3D on the basis of this better baseline.

First of all, take a look at ranging and speed measurement. From end to end, ranging and velocity measurement have obvious advantages on three kinds of dynamic obstacles. The lower the error here, the better, and here is the percentage of error reduction, especially the speed advantage is particularly obvious. Trajectory comparison is a cascade model, its input is the dynamic post-processing of 3D detection on BEV, and some vector information of lane line static post-processing after environmental reconstruction, which is a separate trajectory prediction model. The end-to-end input is the image, and the direct output is the trajectory. Compared with the two, end-to-end has obvious advantages in the position error of this point series.

Here is also a visual video. The box in the picture is the target, the yellow line represents the horizontal and vertical speed, and the pink line is the trajectory. It can be seen that the target coming from the right can appear stably when it passes in front, although it is blocked by the target here. For the speed advantage mentioned just now, especially for the target parked next to it, if it is a detection frame, once the jitter is beyond the range of post-processing energy cover, there will be a relatively large speed, which will trigger a trajectory and cause some brakes of the car. Like here, parked next to it is very stable, and the speed is basically consistent with reality.

05

Real vehicle deployment and closed-loop verification

The last part introduces several key problems related to the model when the real vehicle is closed-loop.

First of all, the model must be quantified and compiled before deployment. Professional quantitative processes and methods, we also arranged an open class to share later. I mainly look at this matter from the perspective of a user.

For the general model, such as CNN-based, we directly single-task float and directly train qat after training, which basically does not lose points. A freezebn will be added in the middle of multiple tasks, so that the counted bn can be applied to each task.

For complex models, a step calibration will be added. The key point here is calibration. Take our end-to-end model deployment as an example. As far as the single-task model is concerned, there are many drop points in the early end-to-end direct float training qat. After consulting my colleagues in the tool chain, I suggest that we do calibration. However, after calibration, it is found that although qat has been greatly improved, it still has a big gap with float. Colleagues in the tool chain suggest fixing the scale of weight or feature. Actually, it is more effective to fix the scale of the feature, but it does not completely solve the problem of qat output of the above tasks from end to end.

So what’s the problem? I also went to some pits. In fact, it is very simple to think about it now. We need to calculate the appropriate feature scale and select the appropriate data when doing calibration. For example, in the early stage, we used the data of urban scenes to estimate the feature scale of speed, and found that the qat index of the trained qat model in high-speed scenes was worse than that of the float. Later, through the tools of the tool chain, the model was analyzed layer by layer and it was found that the problems on the model could not be found. If there is nothing wrong with the model, it is a problem with the data. Finally, we use a high-speed data set. Because calibration itself has few training steps, it is necessary to select the quantity that can cover the perception and reach the upper and lower limits of the maximum and minimum values, so as to estimate a more suitable quantitative scale.

Despite this step, the problem of quantification has been greatly solved. However, the standard of tool chain is generally that the gap between qat and float cannot be greater than 1%. After these two steps, it still can’t reach this level. So what else went wrong? Later, it was also a groping, and it was found that many quantities of 3D perception on BEV were physically meaningful. The final conclusion is to fix the scale for some physical computing layers. Where does the fixed scale come from? For example, as mentioned earlier, the feature on the query index BEV has warping operation, and it has clear physical significance to calculate the two inputs of offset. When the model structure is determined, the maximum and minimum values of the meshgrid on BEV, as well as the maximum and minimum values of the query index, have been determined, so that the maximum and minimum values of the matrix can be calculated, and then the fixed scale can be calculated.

For example, taking odometry as the model input, the data of calibration may be collected at a medium speed, but the speed of the self-propelled vehicle may also change during the actual large-scale data training. Therefore, the maximum and minimum values can also be obtained for this physical quantity with clear rules. With a similar idea, check some numerical calculation layers with physical significance in the network and calculate the fixed scale. Through these operations, the qat problem of the end-to-end model is solved, and it is also a lot of experience. Let’s share it with you.

Quantization is compilation. Compilation means that the compiler directly converts the intermediate format hbir from qat, and finally gets the hbm format deployed on the real board. The specific details are also shared and introduced later. Here is mainly about the performance of BEV perception in actual deployment.

The large-scale BEV is the perception range of 100 forward, 30 backward and 50 left and right respectively, and the small-scale BEV is the perception range of 20 forward and 10 meters in the other three directions, so as to do the element perception on BEV. On the single core of Journey 5, such as a small parking area, the multi-task model of single core and single frame can reach more than 70fps; The wide range of driving can also reach more than 30fps. Even with end-to-end, the model directly output to trajectory prediction can reach more than 20fps, which is in the case of single core. There are two cores on Journey 5, and the frame rate can be further improved in model scheduling.

At the time of deployment, the generalization of the model is tested. Because the model of the customer’s car is different, the corresponding camera installation position and orientation may be different. Our BEV model, especially in the prototype development stage, has many customers who have been verified in the early stage, so a model can support the deployment of multiple models as much as possible, which requires a high generalization of the model. We have also explored a lot, and finally found that the four-plane spatial fusion method introduced earlier has laid a good foundation at the beginning. The example given here is that the training data are all collected by SUV, and the data of cars are used for evaluation. If it is a single plane, it has more than seven drop points compared with four planes.

On the basis of better generalization of the four planes, we will develop some enhancement strategies for training. There are two effective methods here: one is that the BEV model disturbs the image by the camera rpy in one stage, simulates the bumps in the process of vehicle driving, and enhances the image; In addition, it is assumed that the plane is in the place where the angle of view is converted, and then the plane enhancement is made on the plane where the camera leaves VCS, which is also assumed to be the disturbance of rpy.

With these two disturbance strategies, this performance (AP 72.60) can be achieved in the training and training of SUV models and measured in the car data set, which is similar to the performance of an adaptive car model that we have trained with nearly 1 million data collected in about 20 days. Therefore, these strategic feature ensure the generalization of our current BEV perception model in actual deployment.

After deployment, the speed of closed-loop iteration of the model is tested. Especially in the verification stage of the previous prototype scheme, the demo closed-loop verification of different real vehicle installation and deployment is supported through a total version. For the BEV training framework, any camera configuration can be supported. For example, 11v/9v/7v, 7v may replace the 4v panoramic view with 4v fisheye, or any other multiple perspectives, all of which can participate in the model training. When deploying, compile according to the corresponding camera configuration, and then you can do the corresponding test.

The other is the training speed of the model, especially after the time series fusion. Without optimization, if a single frame is trained for 5 days and 10 frames are linear, it will bring 10 times the training time, which is definitely unacceptable. We definitely need to do optimization, and some students who have developed the framework specially guide us to do optimization. At present, the training speed has increased by 56% after optimization.

Specifically, at present, the multi-task release, multi-task with more than a dozen task heads, and the end-to-end training of time sequence can be completed within one week. The later evaluation or the case CICD of some historical data are all automated, thus ensuring the efficiency of the training and distribution step.

Another important factor, especially in the early stage of real vehicle testing, will feed back some non-working case-bad case data. How to use these data in training? There are generally two kinds of optimization for Badcase: one is to mine from the collected data, which is more conventional; The other is to be able to use the returned Badcase data directly.

For things like 2D perception, the link we did in mass production before was relatively mature. For BEV perception, the difference is to see how the model of the brush library comes from or how the data sent to the target comes from. For dynamics, a pure visual large model is used, which is not limited by on-board op and computational power. We have a big model team, and the pure visual BEV big model specially developed has also been used in the dynamic Badcase processing link to generate a pseudo GT to participate in the training. Static, a purely visual mapping scheme, can also send the base map to the standard, and has been deployed to the static perception Badcase link on BEV. In fact, after the deployment of the model, many of them are infrastructure problems. At present, the infrastructure of horizon’s perception of BEV is basically mature.

That’s all I shared today, mainly to convey a few meanings. First of all, based on the fact that Journey 5 can do many things, we have also done many things on it. For the BEV sensing scheme, we can see that we have a better baseline for reference and iteration, so the conclusion is relatively solid.

In addition, if some online students are developing based on Journey 5, or are about to develop based on Journey 5, when you encounter quantitative problems, don’t give up too early. Besides considering the model structure, analyze the data from the perspective of physical meaning and give it a suitable scale.

Finally, I wish everyone a smooth development, and my sharing today is over. Thank you!

Special feature of 1905 film network "Today is a day of celebration! "Feng Shen" has finally reached 2.5 billion! "September 2, withThe first part of feng ShenBox office exceeded 2.5 billion, actorsThis sandFulfilling the promise has brought fans a long-awaited "meatball head" tutorial.

At the same time, Yu Shi, Hou Wenyuan, Yang Le, Huang Xiyan, Shan Jingyao, Pengyi Han and other leading actors are not to be outdone, and they have brought "unique skills" such as dancing, showing muscles, playing Tai Chi, chasing pigs and washing their hair upside down. The Eight Immortals crossed the sea and showed their magical powers.

When the movie box office achieved good results, it was a routine operation for the founders to give back benefits to fans.However, unlike the previous materials and photos, the first part of "Feng Shen""Put all your eggs in one basket", "Learn from Dad" and "Disappearing She"Wait for a fewsummermovie seasonThe film has reached a new height in the "fancy whole life", which can be said that there is no maximum volume, only more volume.

The first part of feng Shen

The first part of Feng Shen let the audience know a group of new actors with both value and strength in one breath.In order to repay the support of the "tap water", the founders launched the welfare without ambiguity and kept their word.

Huang Xiyan, Naernasi, Shan Jingyao and others all chose to make sports videos at the time of 2.5 billion yuan, from swimming, iron throwing and boxing to Tai Ji Chuan, which fully demonstrated the fruitful results of Fengshen Training Camp.

As a "roadshow victim", Yang Le, who plays Boyikao, also made good on his promise at the time of 2.5 billion, and released photos of abdominal muscles. Eight abdominal muscles are clearly visible, and the image of "Hamburg" is also displayed on the face P, and the terrier is played by himself.

Chen Muchi, Li Yunrui and Hou Wenyuan are more open-minded. When it broke 2 billion yuan, Chen Muchi and Li Yunrui brought God of Wealth shorts dance and down jacket women’s group dance respectively, and the picture was very magical.Even Chen Muchi himself commented: "Three days off the net" and "a person’s life will soon pass".

Hou Wenyuan chose to reproduce the "black history" — — The dance of "Subject 1" will carry out the "collapse of human setup" to the end.

If the above is just a routine operation, then the next few can be said to have taken out the housekeeping skills. Xia YuAt the time of 2 billion, an antique trick was displayed, and two small fish were conjured out of thin air, which was very suitable for Shen Gongbao people.

Yu Shi brought a difficult "dragon sucking water" and a brain-opening "chasing pigs" at 2 billion and 2.5 billion respectively.

To say that the most difficult factor is Lei Zhenzi Pengyi Han’s one-handed handstand shampoo, which shows the super arm strength that ordinary people can’t reach.

Wu Ershan, the director, is the most realistic. First, he released a "proton group" muscle photo, and then officially announced that he would launch a variety show of proton training camp, which is expected to meet the audience as a "welfare" at the box office of 2.6 billion yuan.

Put all your eggs in one basket

The first part of "Feng Shen" is in full swing, and another summer explosion "Put all your eggs in one basket" is also not to be outdone. It is full of fancy, singing and dancing, and it is full of network feeling.

Starring LAY turned into Aladdin’s magic lamp, and gave his fans everything they wanted, from "Little Bear Dance" and "Dinosaurs Against Wolves" to singing the theme song "Thousands of Money Are Gone" with fans, which was a perfect score.

Gina also danced the bear dance regardless of the image. The netizen commented: "I can see that today is a big joy, not Anna."

Zhou Ye and Wang Dalu brought a song "Listen to Mom" (gambling-forbidden version) to the audience and also to Atian in the play.

Sun Yang, who has played a lot of fans, finally brought a Cantonese song at the box office of 3 billion yuan, and made an appointment with fans to "advance to 4 billion yuan"!

Learn from dad

Breaking 500 million yuan of "Learning Dad" can be called the "little dark horse" of the summer file, and the celebration methods of several starring roles are also quite grounded. Bo HuangKeep your promise and jump up the upgrade.Dragon fist, join the snake dance, bobby jump and other fat burning actions, very hard.

Wan Qian brought a piano version of "Up in the Cloud", with guitar and kazudi singing in concert with literature and art, full of fairy tales.

NI YAN, who plays the owner of the group of chickens and dolls, "chicken" herself outside the play, bringing "Liu Zhenzhen’s piano debut". User comments: I don’t have enough skills, and I sincerely come together, as if I saw myself practicing the piano.

Lost in the Stars

No more sincere than the involution is "She Disappeared". Although there are not many special benefits, it is also full of sincerity. NiNiWhen I broke 3 billion, I took 30 selfies in one breath, and compared the gestures of 1 to 30 with my hands, which is called sunbathing the ceiling.

Zhu Yilong carefully prepared the piano version of "Cage", thanking the fans and audience for their support all the way. The Weibo broadcast volume of a single video has exceeded 22 million.

The last summer movie set a new record with a box office of 20.6 billion yuan, which is inseparable from the strong support of China audience.

The fancy animation of the film’s main creators not only shows their other side, but also responds to the voice of netizens, which often plays a good marketing and communication effect in the short video era.

But in the final analysis, these are the icing on the cake after all. Only by taking good pictures and becoming explosions can we have the opportunity to do business in a variety of ways.

Who impressed you the most about the above films? What kind of heavy benefits do you want to see next?

　　Moral education curriculum is the most important carrier of moral education in primary schools. Reviewing the changes of moral education curriculum in primary schools since the founding of New China 70 years ago, and summarizing the experience of moral education curriculum construction will help to better implement the fundamental task of cultivating people by virtue and cultivate socialist builders and successors.

　　Four stages of the development of moral education curriculum in primary schools and their characteristics of the times

　　Before the reform and opening up — —

　　There is no unified moral education curriculum.

　　In March, 1952, the Ministry of Education promulgated the Provisional Regulations for Primary Schools (Draft), pointing out that primary education should "give children a comprehensive basic education", in which the purpose of moral education is to make children have good qualities such as patriotism, national morality, honesty, courage, unity, mutual assistance and discipline, and the main content is the "five loves" education of loving the motherland, people, labor, science and public property. In 1953, the State Council issued the Instruction on Rectifying and Improving Primary School Education, and primary school moral education began to emphasize the education of daily behavior norms. The first primary code, published in 1955, put forward specific requirements for the daily behavior norms of primary school students. Generally speaking, from the founding of New China to 1956, moral education in primary schools focused on the cultivation of moral quality and behavior habits, which was more in line with children’s characteristics. However, with the expansion of the "anti-Rightist" struggle in 1957, this situation has undergone great changes.

　　In April 1958, the Central Committee of the Communist Party of China held a national conference on education. The meeting pointed out that education is a tool of class struggle and should serve politics and production. In September of the same year, the Instructions of the Central Committee of the Communist Party of China and the State Council on Education clearly stated that "the Party’s educational policy is that education serves the politics of the proletariat and combines education with productive labor". Since then, moral education courses have shifted from paying attention to moral education to political education. Jiangsu, Zhejiang and other places took the lead in setting up political courses in primary schools, and then gradually expanded to most parts of the country. At the same time, the Ministry of Education requires primary schools to add a class meeting every week to educate students on current affairs and politics, and the political tendency of moral education in primary schools begins. During the ten-year "Cultural Revolution" that began in 1966, the normal school education order was destroyed and school moral education was completely politicized.

　　Reform and opening up to the end of the 20th century — —

　　From political lessons to ideological and moral lessons

　　Before the reform and opening up, there was no unified national curriculum for moral education in primary schools. In 1978, the Ministry of Education promulgated the "Trial Draft of Full-time Ten-year Primary and Secondary School Teaching Plan", which stipulated that politics classes should be offered in the fourth and fifth grades of primary schools, with two class hours per week, mainly to carry out preliminary communist ideological education and necessary political knowledge education. The unified political course has not changed the political tendency of moral education in primary schools, but it was abolished in 1981, so it can be regarded as a transition. However, it is the first time in New China to explicitly set up a political course in the teaching plan, which is of great institutional significance for moral education to get on the right track.

　　From April 22 to May 7, 1979, the Ministry of Education held a national symposium on ideological and political education in primary and secondary schools. After the symposium, the Minutes of the National Symposium on Ideological and Political Education in Primary and Secondary Schools were issued, which criticized the formalism and adult practices in moral education and pointed out that "the ideological and political work of primary and secondary school students must proceed from reality, pay attention to the age characteristics of teenagers, be targeted and stress practical results". This has played a positive role in correcting the politicization and adult tendency of primary school political lessons. In March 1981, the Ministry of Education promulgated the "Full-time Five-year Primary School Teaching Plan (Revised Draft)". In the "Explanation on Revising the Teaching Plan of Full-time Five-year Primary School", it is pointed out in particular: "At present, the political lessons in the fourth and fifth grades are divorced from the students’ ideological reality, and the effect is not good. According to the spirit of adhering to the Four Cardinal Principles and strengthening ideological education for young people, the current political class will be changed into an ideological and moral class, with one class hour per week for grades one to five. " From the primary school political lesson in 1978 to the primary school ideological and moral lesson in 1981, this important change has effectively corrected the tendency of politicization and adulthood of primary school moral education, and turned primary school moral education into cultivating students’ moral quality.

　　In 1982, the Ministry of Education promulgated the "Teaching Outline of Ideological and Moral Courses for Full-time Five-year Primary Schools (Trial Draft)", stating that the purpose of ideological and moral courses for primary school students is to "make primary school students initially have communist moral quality and good behavior habits, and aspire to be workers with ideals, morality, culture and discipline, so as to lay an ideological foundation for training them to be successors of the proletarian cause". Around this goal, the ideological and moral course takes "five loves" (love the motherland, love the people, love labor, love science and love socialism) as its basic content, and combines with the implementation of primary code to educate primary school students on the moral quality and behavior norms that citizens in socialist countries should have, and strives to arrange the teaching content reasonably according to the age characteristics of primary school students. This is the first time that New China has systematically and scientifically constructed the content system of moral education courses in primary schools. Later, according to the requirements of social development and the needs of the situation, the syllabus of ideological and moral course in primary schools (1986), the syllabus of ideological and moral course in full-time primary schools with nine-year compulsory education (preliminary draft) (1990) and the syllabus of ideological and moral course in full-time primary schools with nine-year compulsory education (for trial implementation) (1992) were promulgated successively. Through the continuous revision of the syllabus, the content of ideological and moral course is more in line with the age characteristics of primary school students, and it is more moral enlightenment and life-oriented for primary school students.

　　In 1994, "Several Opinions of the Central Committee of the Communist Party of China on Further Strengthening and Improving Moral Education in Schools" was promulgated, pointing out that the moral education system in schools should be planned as a whole. To implement this opinion, in 1997, the State Education Commission issued "Curriculum Standards for Ideological and Moral Courses in Primary Schools and Junior High Schools in Nine-year Compulsory Education (Trial)", which is the first time to replace the syllabus with curriculum standards, and to determine the teaching objectives and requirements of moral education courses in primary schools and junior high schools as a whole, plan the teaching content system of ideological and moral courses in primary schools and junior high schools, and complete the task of integrating ideological and moral courses in primary schools with junior high schools.

　　Compared with the previous syllabus, the curriculum standard of primary school ideological and moral course in 1997 emphasized the cultivation of good moral quality and civilized behavior habits, and highlighted the foundation and extensiveness of moral education. In the form, it emphasizes the hierarchy, stage, continuity, step by step and spiral rise of teaching.

　　New Curriculum Reform in the 21st Century — —

　　Offering courses on morality and life, morality and society in stages.

　　In 2000, the general offices of the General Office of the Central Committee of the CPC and the State Council issued the Opinions on Further Strengthening and Improving Moral Education in Primary and Secondary Schools to Adapt to the New Situation, proposing that we should proceed from reality, thoroughly study the current ideological and moral characteristics of students, revise the curriculum standards of ideological and moral courses in primary schools and ideological and political courses in middle schools, further reform and improve the teaching content, and strive to build a moral education curriculum system in primary and secondary schools to meet the needs of development in the 21st century. According to this opinion, the Ministry of Education made necessary revisions to the curriculum standards of ideological and moral education in 1997, and in 2001, it promulgated the Curriculum Standards of Ideological and Moral Education in Primary Schools and Ideological and Political Education in Junior High Schools for Nine-year Compulsory Education (Revised). However, with the start of the new curriculum reform in 2001, the ideological and moral course has been changed into moral and life, moral and society, with new curriculum standards. In 2001, the curriculum standards of ideological and moral education in primary schools only played a transitional role.

　　In June 2001, the Ministry of Education promulgated the Outline of Basic Education Curriculum Reform (Trial). The "Outline" proposes "to set up a nine-year compulsory education curriculum as a whole". In primary school, comprehensive courses are the main courses, and morality and life are offered in lower grades, while morality and society are offered in middle and high grades. In 2002, the Ministry of Education officially promulgated the Curriculum Standards of Morality and Life in Full-time Compulsory Education (experimental draft) and the Curriculum Standards of Morality and Society in Full-time Compulsory Education (experimental draft).

　　From ideological and moral course to moral and life, moral and society, these two course standards and ideological and moral course standards have changed greatly in course nature and design ideas. The changed course name emphasizes activity and comprehensiveness more, which is more suitable for the characteristics of primary school students in low and middle grades. Moral and life curriculum is connected with kindergarten curriculum, which is an active comprehensive curriculum based on children’s life and aimed at cultivating children with good moral character, willingness to explore and love life. It is designed according to three axes: children and self, children and society, and children and nature. Morality and Society course is a comprehensive course based on children’s social life and promoting students’ good moral character formation and social development. It focuses on children’s social life, from individual, family, school to community, hometown, motherland and the world, and forms six modules, such as "I am growing", "I am with my family", "I am with my school", "I am Chinese" and "approaching the world". In 2011, the state promulgated the curriculum standards of Morality and Life, Morality and Society (2011 edition), which kept pace with the times in content and reflected the requirements of the times. At this stage, the management of teaching materials is "one syllabus and many books", which is examined and approved by the textbook review Committee of primary and secondary schools of the Ministry of Education.Multiple sets of teaching materials are selected for use at the same time.

　　Entering a new era — —

　　Use the unified textbook "Morality and Rule of Law"

　　On October 23, 2014, the Fourth Plenary Session of the 18th CPC Central Committee adopted the Decision of the Central Committee of the Communist Party of China on Several Major Issues Concerning Comprehensively Advancing the Rule of Law, which proposed that the rule of law education should be integrated into the national education system, and the rule of law knowledge courses should be set up in primary and secondary schools. To implement this Decision, in June, 2016, the Ministry of Education, the Ministry of Justice and the National Law Popularization Office jointly issued the Outline of Youth’s Legal Education, requiring timely and corresponding revision of moral education curriculum standards in primary and secondary schools to complete the teaching content required by the outline. Starting from the fall semester of 2016, the state has changed Morality and Life, Morality and Society in primary schools and Ideology and Morality in junior high schools into Morality and Rule of Law, and set up special books on the rule of law in the second volume of grade six and the first volume of grade eight.

　　On the one hand, the compilation of the textbook Morality and Rule of Law for primary schools is based on the curriculum standards of Morality and Life, Morality and Society and the Outline of Education for Teenagers under the Rule of Law; on the other hand, it is based on the fundamental task of cultivating people in the new era, highlighting the guidance of socialist core values, and strengthening the education of Chinese excellent traditional culture, revolutionary culture, rule of law, national sovereignty awareness and national unity. "Morality and Rule of Law" is compiled by the Ministry of Education, with "one syllabus and one book", reviewed by the National Textbook Committee and used nationwide. Since 2017, the unified textbooks have been used in the first and seventh grades, in the second and eighth grades in 2018, and in the autumn of 2019, all grades will be fully covered. In January 2019, the Ministry of Education initiated the revision of the curriculum standards for compulsory education, and the curriculum standards for Morality and Rule of Law are currently being formulated.

　　Future-oriented,Constructing the overall moral education curriculum suitable for children

　　Since the founding of New China 70 years ago, moral education courses in primary schools have grown from scratch, from political lessons to ideological and moral lessons, from ideological and moral lessons to morality and life, morality and society, and then to morality and rule of law in the new era. The course has both a changing side and a constant side. What remains unchanged is the initial intention of moral education curriculum — — To cultivate all-round socialist builders and successors, the contents and organizational forms of moral education have changed, such as from a single political education to paying attention to moral education and then to both morality and law, from instilling political dogma to constructing children’s moral life, from meeting the requirements of society to meeting the common needs of social development and individual development, from "one program with many books", choosing to use one program with one book, and using it in a unified way.

　　Looking forward to the new era, the development of moral education curriculum in primary schools should achieve four points: first, put the moral education curriculum in primary schools into the overall framework of "the integration of moral education in colleges, middle schools and primary schools" and design the moral education curriculum in primary schools step by step and spirally. Second, take children as the center, construct children’s moral life, including children’s communication life with themselves, others, society, country and the world, transcend individuals and realize the integration of individuals and society. Third, construct the content system of "Great Moral Education" course. Moral education is not only a political course, but also an ideological and moral course. It is a "grand moral education" that includes ideological and political education, moral education, rule of law education, psychological education and traditional culture education. The content of moral education not only has a stable side, but also keeps pace with the times and embodies the times. Fourth, highlight the national will of moral education curriculum. General Secretary of the Supreme Leader pointed out: "Ideological and political course is the key course to implement the fundamental task of Lide Shuren." In the new era, we must adhere to the Party’s overall leadership over the construction of ideological and political courses, make unremitting efforts to cultivate people with Socialism with Chinese characteristics Thought, the supreme leader of the new era, strive to cultivate new people of the times who are responsible for national rejuvenation, and cultivate socialist builders and successors who are fully developed morally, intellectually, physically and artistically.

　　(The author is distinguished professor, director of the Institute of Moral Education of Nanjing Normal University and Changjiang Scholar of the Ministry of Education)

Welcome to contribute | Please add WeChat shaoyanlin412.

According to the survey report on the living conditions of national automobile dealers in 2023 released by china automobile dealers association, more than 70% of dealers failed to complete the annual task index; The loss ratio accounts for 43.5%, and the profit ratio is only 37.6%.

Will car dealers and 4S shops still be a good business in 2024?

Let’s take a look at several sets of key data. (The following data source: china automobile dealers association)

According to the survey, 27.3% of the vehicle sales targets are achieved at or above 100%, and 19.4% are between 90% and 100%, which make up 46.7%. Usually, the completion rate is above 90%, and the OEM can get 100% rebate.

This means that in 2023, more than half of dealers failed to get the "planned" rebate at the beginning of the year.

Judging from the proportion of income of each business module in the total income, the proportion of new cars and after-sales has not changed much, only the proportion of financial insurance has been rising all the way, and the proportion of used cars has been declining all the way.

The profitability of dealers is also not optimistic. In 2023, the loss ratio reached 43.5%, and the profit ratio was 37.6%. Although the profitability improved compared with that in 2022, it was significantly lower than that in 2021, which was the second worst result in the past six years.

Finally, look at the composition of dealers’ profits. The vehicle sales contributed -15.6% and the after-sales contributed 60.7%. Obviously, the new car does not make money, and the after-sales volume is small, which is the dilemma faced by dealers.

Combining these tables, we can roughly see what kind of year the dealers experienced in 2023.

First of all, dealers are trapped in the traditional business model.None of the business sectors can make people shine at the moment, leading a certain sector to rise suddenly.

Vehicle sales always account for about 80% of operating income, which confirms an experience in the industry: vehicle sales determine whether dealers are dead or alive.

Yes, after-sales can’t determine the life and death of dealers. Even those dealers whose service absorption rate is more than 100% will actually live for less than three years after being cancelled and turned into ordinary repair shops.

Because on the one hand, there is no source of supplementary customers, and the loss of customers and the reduction of admission are inevitable results; On the other hand, the expenses such as rent, water and electricity personnel’s wages have not decreased, and the profits will only become thinner and thinner; The most important thing is to lose the gold-lettered signboard authorized by the manufacturer, and the degree of trust is no longer the same.

Secondly, new car sales not only don’t make money,It is also necessary to dig out a piece from the profits of other business sectors to fill the "negative profits".

Combined with the achievement rate of vehicle sales target, the dealers who can make profits or survive in 2023 are all the stores that rank higher in the national sales ranking of brands or top in the regional sales, with high achievement rate. They can not only get millions of full rebates, but also get hundreds of thousands of extra rewards such as "market contribution award" and "market development award".

From the OEM’s point of view, if all the head dealers in the brand fall down, the brand will have no one to follow. In order to prove the "correctness" of its marketing strategy, it is necessary to set a good example, even if we put some water on these dealers at a critical moment.

From the dealer’s point of view, only those senior dealers whose service absorption rate is greater than 100% can have the confidence to reduce the profit of the new car sales section in the store to negative, so as to follow up the annual volume price in 2023 and sell cars at a loss. The gambling is that the main engine factory does not dare to let itself go bankrupt.

Especially for those dealers whose land is owned rather than leased, the operating expenses are missing a large piece of land or store rent, and only the personnel expenses and marketing expenses are left in the big expenses.

Even if the profits of previous years are filled in, the speed of losing blood can be greatly alleviated, and they can stay at the poker table for a longer time, and the greater the grasp of procrastination.

So in 2024, is automobile distribution still a business project worth sticking to?

Needless to say, the enterprises or 4S shops that still have the status of car dealers are all authorized to sell fuel car brands (except Tucki).

Whether these dealers are still worth sticking to in this industry depends on the power technology route they will adopt in the future and their geographical location.

With regard to the future development direction of automobile power technology route, pure electric vehicles will become more and more mainstream, but it will not be realized in a short time, and its market penetration may even fluctuate upward, let alone completely replace fuel vehicles.

Even if the hybrid vehicle is a transitional technology, this transitional period will last at least 10 years, or even more than 20 years.

Fuel trucks or liquid fuel trucks will not disappear completely. Even in the southern provinces and cities of China, there will still be sporadic use in emergency rescue and disaster relief, field work, etc., while the scope and time of use in the north will be wider and longer.

The power technology route of automobile brand OEM is a long-term strategic decision that will not be easily changed. As its distributor, it is relatively easy to obtain accurate information.

Then the remaining major influencing factor is the geographical location of the dealer.

If the dealer is located in the north of Qinhuai line, even if the power technology route of the brand main engine factory is only fuel oil and electricity, the risk of sticking to it is still relatively low; If the power technology route of the brand OEM is only pure electricity, then the risk of sticking to it is relatively high.

On the other hand, if the dealer is located in the south of the Yangtze River, the risk of sticking to technology brands with only fuel and oil is relatively high, even if the risk of sticking to technology brands with only pure electricity is still relatively low.

So, in 2024, is automobile distribution still a business project worth investing in?

This problem is more complicated than the previous one, which involves not only the power technology route that the brand OEM will adopt in the future and the geographical location of the proposed store, but also the business policy of the brand OEM.

Because the brand OEM sells well and makes money or not, it doesn’t mean that the dealer sells well and makes money or not.

At present, most of the "new power" new energy brand OEMs adopt a mixed cooperation mode, that is, the electromechanical service sector of new car sales and after-sales is self-operated, and only the after-sales car body panel is recruited.

Typical brand OEMs, such as Tesla (since last year, it has started to operate its own business), Weilai, Ideality, etc., are not discussed here.

Most of the new energy brands established by the original fuel vehicle brand OEM still choose the 4S store model, such as:, Lantu, Feifan, Zhiji, Aouita, etc., but when choosing these brands to apply, we must pay attention to their respective business policies.

In business policies, special attention should be paid to the following terms:

Address of the proposed store:Generally, it is required to be in the automobile business circle or downtown area, which is related to the rental cost in future operations.

Building area and building form:It is related to the capital investment of hardware construction and the cost of hydropower publicity in operation.

Sales commission:New energy brands basically adopt the sales agency system, that is, dealers are no longer required to control the inventory, but dealers also lose their influence on the terminal retail price of automobiles. Every time a commercial car is sold (signed), they can only get a sales service fee of 2%-3% of the sales price of new cars from the brand OEM, that is, a sales commission.

This profit is basically less than half of the difference between the purchase and sale of dealers 10 years ago, but for the current market environment, it is a long-lost new car sales profit. However, it is unclear how the proportion of this sales service fee will change in the future.

Payment method of store building rebate:Brand OEMs usually give rebates to newly-built stores, but the current rebate method is disgusting. For example, the total amount of the promised rebate for building stores is 2 million, but it needs to be paid in three years (every 12 months is one year from the month after acceptance and opening), with a maximum of 700,000 yuan per year, which is paid as a subsidy to 1000 yuan for each new car sold.

That is to say, in the past three years, at least 700 vehicles must be sold every year, in theory, in order to get the store-building rebate in full. After the new store opened for three years, the sales volume usually increased gradually. If 500 units, 800 units and 1,000 units were sold in these three years, the store-building rebate would be 50+70+70 = 1.9 million. Of course, in order to reach the sales volume of 700 units in the first year, investors increased their investment in publicity or marketing activities to push up sales volume, and brand OEMs also encouraged and welcomed the success, because it was the investors’ money that was spent, and these promotional investments, brand OEMs said that even if it rains, it will never rain.

In short, there are many reward or rebate clauses in business policies that need to be carefully studied. If you want to get them in full or at the top, there is basically only theoretical feasibility, just for the operation of addition, subtraction, multiplication and division, which is quite insulting to IQ.

As for whether it is worth investing, it is necessary to judge the specific location, brand and investor’s own conditions one by one.

Article source: automobile service world, thank you.

　　"Teacher Lu playing with bones" Lu Jing put the bones of fish head, braised chicken and Beijing roast duck together into a whole skeleton, creating a kind of "visual sense of paleontology museum".

　　Dai Jianye, a professor of "Debris Flow", told the story of Li Bai, Du Fu and Gao Shi joining hands to visit famous mountains, visit immortals and refine the elixir, which reappeared the prosperous time of the Tang Dynasty.

　　Gong Gongjun, who only made a noise but didn’t show up, took everyone to visit every corner of the Forbidden City and quoted the classics to tell the cold knowledge and little interesting things of the Forbidden City.

　　With the development of short video platform, more and more content creators have emerged to share and spread knowledge. They are popular science experts who have knowledge, love sharing and are familiar with skills. These high-quality content supplies meet and create a huge demand for knowledge.

　　Professor Hu Baijing, Executive Dean of School of Journalism, Renmin University of China, believes that the rise of short video platform has promoted the expansion of knowledge production from elite to the public, making knowledge more personalized and scene-oriented, and at the same time making it possible for knowledge to be universal, knowledge to be promoted, and knowledge to be shared and co-created.

　　

　　Small but refined, making universal knowledge close to reality.

　　"Tang Xuanzong released Li Bai from the court, that is, he was given a sum of money to get rid of him: you are not cut out to be a prime minister. Li Bai went to Luoyang from Chang ‘an, and Du Fu was a fan of Li Bai. Li Bai was kind enough to let him meet him. After meeting him, he admired Li Bai to death. I don’t know whether Li Bai plays cowhide well or is charming. Anyway, Du Fu listened to his words and started from Luoyang, Henan Province to Hebei and Shandong with his eldest brother. What for? Looking for immortals, picking fairy grass and refining elixir, I met Gao Shi halfway … "This is the story that happened in 744 AD and was called the meeting of the sun and the moon in China literary world by Wen Yiduo. Only Confucius can match it when he saw Lao Zi.

　　Although it doesn’t sound formal, this remark is by no means a fabrication by Dai Jianye, a professor at the College of Literature of Central China Normal University. The experiences of Li, Du and Gao in seeking immortality and monasticism are reflected in several poems of Du Fu. Dai Jianye lamented that it was a generation of romantic people.

　　When this lecture video was put on Tik Tok, Dai Jianye didn’t know what the latter was. More than a year later, he became Professor online celebrity known as the Debris Flow. The Tik Tok has nearly 4 million fans and tens of millions of likes.

　　The "big family" living in Hefei is a lawyer with nearly 10 years of financial work experience. In her spare time, she uses short videos to give netizens common knowledge of popular science and economics.

　　The "treasure family" found that the public is very curious about knowledge, but many people have no chance to receive economic education, lack basic knowledge of wealth, and have no concept of risks and benefits. If the public can learn more about economics, they may be able to avoid some losses. Some netizens left a message under the short video she posted: "Fortunately, I saw that you were not fooled by the Ponzi scheme, and my friends around me pressed hundreds of thousands of yuan on a financial platform."

　　In the era of short video, knowledge levels are diversified and symbiotic, including authoritative scientific theories, life tips and workplace office skills to meet the different knowledge needs of different audiences. The sharers of knowledge are not necessarily high-ranking experts and scholars, but also life experts who are familiar with their own fields.

　　Compared with the traditional graphic form, short video communication knowledge has four characteristics: immediacy of knowledge dissemination, personalization of knowledge presentation, manifestation of invisible knowledge and popularization of complex knowledge.

　　Zhao Yan, President and Editor-in-Chief of China Science Newspaper, said that a short video of 15 seconds to 1 minute condensed the serious knowledge on weekdays and passed it on to the audience in a more visual form, which changed the boring appearance of knowledge and narrowed the distance between the public and cultural knowledge. Knowledge sharing and linking with short videos not only broadens the boundaries of knowledge, but also brings knowledge inclusiveness close to reality.

　　Promote the value shift of Internet content industry.　

　　In the eyes of knowledge creators, short videos have become the primary tool for them to spread and exchange knowledge.

　　Xiang Bo, who works as a chemistry teacher at Guangyuan Middle School in Sichuan Province, said that his own class could only affect 40 or 50 students in the classroom. "But now I am a teacher shared by more than 6 million fans. Making good use of communication tools can eliminate the time and space barriers of knowledge dissemination and light up more people’s enthusiasm for chemistry. "

　　Lu Jing, an associate researcher at the Institute of vertebrate paleontology and Paleoanthropology, Chinese Academy of Sciences, feels that doing popular science on short videos is a "new and wonderful experience". Before that, she didn’t expect that the mobile internet platform could tilt resources to serious knowledge creators and content like her. Short videos made her find a hyperlink to share knowledge with the world, and the feedback from "fans" became her motivation to persist in doing popular science. "I feel that short videos make the way to do knowledge on the Internet wider and wider."

　　The purpose of people’s initial use of short video software is mostly entertainment and leisure, but in a highly fragmented and entertaining viewing environment, it is easy to have a sense of emptiness and aesthetic fatigue after a long time. The emergence of knowledge-based content, like a clean stream, can fill the gaps in people’s hearts.

　　Wang Xiaohong, a professor at Communication University of China, pointed out that with some Internet platforms successively launching measures to spread service knowledge, "knowledge content has gained unprecedented attention in the Internet industry. This trend reflects the upgrading of Internet content from single entertainment to knowledge. It can be said that the short video platform has driven the value shift of the Internet content industry. "

　　For users, short videos break the original time-space limitation of knowledge intake, allowing people to use their spare time to study anytime and anywhere. At the same time, short videos present knowledge in the form of concise and clear knowledge points, and ignite users’ interest in further in-depth learning.

　　Therefore, on the one hand, short video greatly reduces the threshold and difficulty of knowledge reception, mobilizes the enthusiasm of the public to create and disseminate knowledge, and realizes co-creation; On the other hand, short videos take social interaction as a link, breaking down the barriers of knowledge dissemination, allowing knowledge to reach more people and realize sharing.

　　Provide a better environment for knowledge content creation

　　"Information creates value. Among the many values created by information, the value of knowledge dissemination is the highest." Zhang Yu, vice president of ByteDance, believes that continuous investment in knowledge content will benefit the sustained and healthy development of the entire Internet content industry.

　　"As a platform or technology, if high-quality content is not occupied, then harmful content will be occupied." Dai Jianye said, "Short videos have both pictures and sounds, and you can also add words, which is definitely a very good way to spread. Algorithms and traffic tell us what kind of articles and videos are the most popular and what kind of knowledge people like, which not only enables the author to tailor them accurately, but also enables the network to deliver them to the door regularly and quantitatively, which profoundly affects the production of knowledge. Those popular knowledge will be churned out in large quantities, but the neglected knowledge will not be produced. "

　　Dai Jianye’s viewpoint is being verified by practice. At the "2019DOU Knowledge Creator Conference" held in Beijing a few days ago, Zhang Yu introduced that as of December 2, 2019, the number of knowledge content creators with over 10,000 fans in Tik Tok had exceeded 74,000, and a total of 19.85 million high-quality knowledge short videos were created, with a cumulative broadcast volume of more than 1.9 trillion; Each short video of knowledge has reached nearly 100,000 people.

　　At present, all major short video platforms have built their own content pools, focusing on providing young people with high-quality content such as teaching, calligraphy and painting, parenting education, human history, traditional culture, handcrafting, natural science popularization and so on. For example, the Encyclopedia, a knowledge label area initiated by Aauto Quicker, and the DOU Knowledge Plan and the Non-legacy Partner Plan initiated by Tik Tok in conjunction with the Science Communication Bureau of Chinese Academy of Sciences, etc.

　　However, at present, the overall proportion of these contents is still low, and the total amount is still insufficient. In the future, various short video platforms will continue to exert their efforts to provide a better environment and more support for knowledge content creation.

　　In addition, regarding the view that short video learning is fragmented, Chen Zheng, a teacher at the National Physics Experimental Teaching Demonstration Center of Beijing Jiaotong University, said in an interview with the media: "Short video really can’t achieve the effect of systematic teaching, nor can it train professionals, but at least it can do one thing, that is, open a door. I teach physics, and my goal is to show some basic world outlook and methodology to the audience, especially teenagers, through these fragmented puzzles. When they collect enough’ fragments’, they master their internal connections and can spell out a complete picture of the world. "

Continue reading »