WO2021237768A1

WO2021237768A1 - Data-driven-based system for implementing automatic iteration of prediction model

Info

Publication number: WO2021237768A1
Application number: PCT/CN2020/094133
Authority: WO
Inventors: 董维山; 张驰; 杨文娟; 蒋竺希
Original assignee: 初速度（苏州）科技有限公司
Priority date: 2020-05-29
Filing date: 2020-06-03
Publication date: 2021-12-02
Also published as: DE112020003091T5; CN113799793B; CN113799793A

Abstract

A data-driven-based system for implementing the automatic iteration of a prediction model, comprising a processor and a cloud. The processor labels behavior data of each road participant on the basis of an observed movement trajectory of each road participant to obtain corresponding labeling information, thus achieving automatic labeling, and then sends filtered first behavior data and the corresponding labeling information to the cloud. When determining that preset automatic trigger conditions are met, the cloud extracts training samples to train an initial network model and obtain a target network model, so that model training is automatically triggered, and then evaluation is automatically triggered. Moreover, when evaluation results meet model update requirements, the cloud sends the target network model to the processor to achieve the automatic deployment of the target network model. The processor automatically updates the prediction model to the target network model after receiving the target network model. Hence, no manual participation is required for data labeling, model training, model evaluation or model updating, and the degree of automation is high.

Description

A data-driven system for automatic iteration of predictive models

Technical field

The present invention relates to the technical field of intelligent driving, in particular to a data-driven system for realizing automatic iteration of prediction models.

Background technique

In the autonomous driving scenario, knowing the trajectory of road participants in advance is conducive to their own safe driving. For example, in the autonomous driving scenario of unmanned vehicles, the future trajectory of road participants is determined by the predictive model set on the unmanned vehicle To make predictions, as the complexity of the road scene continues to increase, the prediction model needs to be updated from time to time.

The current method of updating the prediction model is: the developer performs offline annotation on the behavior data of a large number of road participants to obtain the annotation information, and then the developer conducts model training and evaluation based on the behavior data of the road participants and the corresponding annotation information to obtain new predictions Model, and then the developer will update the new predictive model to the unmanned vehicle through network transmission or hard disk connection.

It can be seen that the above-mentioned update method of the prediction model requires manual data annotation and manual triggering of model training and model update, which is highly dependent on humans, resulting in high labor costs and low automation.

Summary of the invention

The invention provides a data-driven system for realizing automatic iteration of prediction models without manual participation, greatly reducing labor costs, and having a high degree of automation. The specific technical solution is as follows.

In the first aspect, the present invention provides a data-driven system for automatic iteration of predictive models. The system includes a processor and a cloud. The processor is provided with a predictive model, and the predictive model is used to predict road participants. Trajectory of future movement;

The processor obtains the behavior data of each road participant, where the behavior data includes the environmental static map information at the current moment and the historical movement trajectory of the road participant before the current moment collected by the collection device installed in itself, based on The motion trajectory of each road participant observed by its own sensor is labeled with the behavior data of each road participant to obtain the corresponding labeling information, and the behavior data of each road participant is selected from the behavior data of each road participant that meets the preset screening requirements First behavior data, sending the first behavior data and corresponding annotation information to the cloud;

The cloud stores the received first behavior data and corresponding annotation information in the original database, and performs feature extraction on the first behavior data according to a preset feature extraction method to obtain a feature extraction amount, and the feature extraction amount is And the corresponding labeling information are stored as training samples in the training sample library, and when the preset automatic trigger conditions are met, the training samples stored in the training sample library after the last training sample is extracted are extracted to train the initial network model to obtain the target A network model, wherein the target network model is used to correlate the behavior data of road participants with the corresponding future motion trajectory, and evaluate the target network model according to a preset evaluation method to obtain the evaluation result, when the evaluation result When the model update requirement is met, sending the target network model to the processor;

The processor receives the target network model, and updates the prediction model to the target network model.

Optionally, the processor uses the motion trajectory of each road participant observed by its own sensor as label information corresponding to the behavior data of each road participant.

Optionally, the processor predicts and obtains the future motion trajectory corresponding to the behavior data of each road participant based on the prediction model, and for each road participant, calculates the future motion trajectory corresponding to the behavior data of the road participant and The difference between the movement trajectories of the road participants observed by its own sensor, and the behavior data of the road participants whose difference is greater than the preset difference is used as the first behavior data;

or,

The processor determines the behavior category corresponding to the behavior data of each road participant according to the label information corresponding to the behavior data of each road participant, and uses the behavior data of the road participants whose behavior category is a preset category as the first One behavioral data;

or,

The processor determines the type of each road participant, and uses the behavior data of the road participant whose type is a preset type as the first behavior data.

Optionally, the preset category is lane changing behavior or overtaking behavior.

Optionally, the preset type is a large vehicle, a pedestrian, or a two-wheeled vehicle.

Optionally, the processor stores the first behavior data and corresponding annotation information before sending the first behavior data and corresponding annotation information to the cloud.

Optionally, when the number of training samples stored in the training sample library after the last training sample is extracted reaches a preset number threshold, the cloud extracts the training samples stored in the training sample library after the last training sample is extracted The training samples train the initial network model to obtain the target network model;

or,

When the time between the last time the training sample was extracted and the current time reaches the preset time length, the cloud extracts the training samples stored in the training sample library after the last time the training samples were extracted to train the initial network model to obtain the target Network model.

Optionally, the cloud predicts the behavior data of each road participant to be predicted in the data set to be predicted based on the target network model to obtain the corresponding future motion trajectory;

Calculate the size of the evaluation index of the target network model according to the future motion trajectory corresponding to the behavior data of each road participant to be predicted and the motion trajectory of each road participant to be predicted observed by its own sensor;

Calculating an increase amount of the evaluation index of the target network model relative to the evaluation index of the prediction model;

When the amount of increase meets a preset increase requirement, the target network model is sent to the processor.

Optionally, the road participants include vehicles and/or pedestrians.

It can be seen from the foregoing that, in the embodiment of the present invention, the processor marks the behavior data of each road participant based on the motion trajectory of each road participant observed by its own sensor to obtain the corresponding label information, and realizes automatic labeling. Annotate information instead of manual offline annotation, and then send the filtered first behavior data and corresponding annotation information to the cloud. When the cloud determines that the preset automatic trigger condition is met, it will extract the training sample library after the last training sample was extracted The stored training samples train the initial network model to obtain the target network model, which can automatically trigger the model training when the preset automatic trigger conditions are met, and then automatically trigger the evaluation after the target network model is obtained, and the evaluation result meets the model When updating the request, the target network model is sent to the processor to achieve the automatic deployment of the target network model. Finally, after the processor receives the target network model, it can automatically update the prediction model to the target network model. In summary, it can be seen that whether it is Data labeling, model training, model evaluation, or model update do not require manual participation, which greatly reduces labor costs and has a high degree of automation. Of course, implementing any product or method of the present invention does not necessarily need to achieve all the advantages described above at the same time.

The innovative points of the embodiments of the present invention include:

1. The processor labels the behavior data of each road participant to obtain corresponding labeling information based on the motion trajectory of each road participant observed by its own sensor, and realizes automatic labeling to obtain the labeling information instead of manual offline labeling.

2. When the preset automatic trigger conditions are met, the training samples stored in the training sample library after the last training sample was extracted can be used to train the initial network model to obtain the target network model. Therefore, the embodiment of the present invention satisfies the preset When the condition is automatically triggered, the model training can be automatically triggered without manual participation, which greatly reduces labor costs and has a high degree of automation.

3. After the target network model is obtained, the evaluation is automatically triggered, and the increase in the evaluation index of the target network model relative to the evaluation index of the prediction model meets the preset increase requirement, that is, the performance of the target network model is better than the prediction model to be updated When the target network model is sent to the processor, the automatic deployment of the target network model is achieved without manual participation, which greatly reduces labor costs and has a high degree of automation.

4. After the processor receives the target network model, it can automatically update the prediction model to the target network model without manual intervention, which greatly reduces labor costs and has a high degree of automation.

5. According to an embodiment of the present invention, a data-driven system for automatic iteration of prediction models is provided. Only one developer can complete the automatic update of the prediction model, which greatly improves the efficiency of research and development and reduces the cost of research and development. .

6. After the target network model is obtained, the evaluation is automatically triggered, and the increase in the evaluation index of the target network model relative to the evaluation index of the prediction model meets the preset increase requirement, that is, the performance of the target network model is better than the prediction model to be updated When the target network model is sent to the processor, the automatic deployment of the target network model is achieved without manual participation, which greatly reduces labor costs and has a high degree of automation.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

FIG. 1 is a schematic structural diagram of a data-driven system for realizing automatic iteration of predictive models according to an embodiment of the present invention.

In Figure 1, 10 processors, 20 clouds.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

It should be noted that the terms "including" and "having" in the embodiments of the present invention and the drawings and any variations thereof are intended to cover non-exclusive inclusions. For example, the process, method, system, product, or device that contains a series of steps or units is not limited to the listed steps or units, but optionally also includes unlisted steps or units, or optionally also includes Other steps or units inherent in these processes, methods, products or equipment.

The embodiment of the present invention discloses a data-driven system for realizing automatic iteration of a prediction model, which can automatically update the prediction model without manual participation, greatly reduces labor costs, and has a high degree of automation. The embodiments of the present invention will be described in detail below.

FIG. 1 is a schematic structural diagram of a data-driven system for realizing automatic iteration of predictive models according to an embodiment of the present invention. 1, a data-driven system for realizing automatic iteration of prediction models provided by an embodiment of the present invention includes a processor 10 and a cloud 20, the processor 10 and the cloud 20 are in communication connection, and the processor 10 is provided with a prediction model and a prediction model. It is used to predict the future trajectory of road participants, where road participants include vehicles and/or pedestrians.

In the embodiment of the present invention, in order to improve the degree of automation, the processor 10 automatically acquires the behavior data of each road participant and makes online annotations. The behavior data includes the environmental static map information at the current moment and the data collected by the collection equipment installed in itself. The historical movement trajectory of road participants before the current moment.

In an implementation manner, the processor 10 can obtain the behavior data of each road participant as follows: it is equipped with collection equipment and positioning system, such as a camera and a global positioning system, so that it can perform real-time monitoring during driving. The collection range is collected and its own location is located. Based on the real-time collected images and its own real-time location information, the current environmental static map information and the movement trajectory of each road participant can be obtained. The environmental static map information can include Road marking information and road sign information, such as: zebra crossing and red street light information.

In another implementation manner, the way that the processor 10 obtains the behavior data of each road participant may be as follows: it is installed with a collection device, a positioning system, and a high-precision map, such as a camera, a global positioning system, and a high-precision map. , It can collect the collection range and locate its position in real time while driving. Based on the real-time collected image, its real-time location information and high-precision map, it can get the environment static map information at the current moment and the participation of each road The trajectory of the person.

Because the way to label the behavior data of road participants is to use the observed real movement trajectory for labeling, and with the passage of time, its own sensor can naturally observe the real movement trajectory of each road participant, that is, In other words, the prediction result given by the prediction model at the current moment can directly observe the true value in the future, that is, it can be observed at a very low cost whether the prediction result of its prediction actually occurs, whether it is correct, and to what extent it is correct. For example: it is an unmanned vehicle and its own sensor is a vehicle sensor. When the prediction model uses the known static map information of the environment at the current time t0 and the historical movement trajectories of road participants accumulated and observed before t0, a certain adjacent lane is made. It is predicted that the vehicle will move to 5 meters directly in front of the vehicle at time t1 in the future. As time goes by, at time t1, the vehicle sensor can directly observe the true trajectory of a vehicle in the adjacent lane. At this time, the behavior data of each road participant can be marked based on the motion trajectory of each road participant observed by its own sensor to obtain the corresponding marking information. The observed motion trajectory of each road participant is used as the label information corresponding to the behavior data of each road participant.

It should be noted that a data-driven system for realizing automatic iteration of predictive models provided by the embodiments of the present invention can be applied to the field of unmanned vehicles, robotics, and other fields that can realize automatic driving. When applied to the field of unmanned vehicles, At this time, the aforementioned processor 10 may be a vehicle-mounted processor, the collection device installed in the own vehicle may be a collection device of the own vehicle, and the own sensor may be a vehicle sensor of the own vehicle.

Due to the limited computing power of the processor 10, after labeling the behavior data of road participants, the behavior data and the corresponding labeling information need to be sent to the cloud for processing, and because not every road participant’s behavior data is It is valuable for model training, so the data that is valuable for model training can be selected from the behavior data of road participants for processing, that is, the first behavior that meets the preset screening requirements is selected from the behavior data of each road participant Then, the first behavior data and the corresponding annotation information are sent to the cloud 20. Exemplarily, the first behavior data can be filtered through a data filter.

In order to avoid data loss, the processor 10 may also store the first behavior data and corresponding annotation information before sending the first behavior data and corresponding annotation information to the cloud 20.

Among them, there are many ways to filter out the first-line data, including but not limited to the following:

The first:

The processor 10 predicts and obtains the future motion trajectory corresponding to the behavior data of each road participant based on the predictive model, and calculates the future motion trajectory corresponding to the behavior data of the road participant and the one observed by its own sensor for each road participant. For the difference between the movement trajectories of the road participants, the behavior data of the road participants whose difference is greater than the preset difference is used as the first behavior data.

Since there is a big difference between the future motion trajectory of a certain road participant predicted by the prediction model and the observed motion trajectory of the road participant, it means that the prediction model is not yet capable of the future motion of the road participant. The motion trajectory is predicted more accurately. Therefore, the behavior data of these road participants with large differences is valuable data for model training. The behavior data of these road participants with large differences can be used to train a new prediction model. When the training is completed, the new prediction model It is possible to accurately predict the future trajectories of these road participants with large differences.

Therefore, when the processor 10 screens the first behavior data, it can predict the future motion trajectory corresponding to the behavior data of each road participant based on the prediction model, and then compare it with the observed motion trajectory, and take the larger difference as The first behavior data, that is, for each road participant, the difference between the future motion trajectory corresponding to the behavior data of the road participant and the motion trajectory of the road participant observed by its own sensor is calculated, and the difference is greater than expected Let the behavior data of different road participants be the first behavior data.

Therefore, for each road participant, the difference between the future motion trajectory corresponding to the behavior data of the road participant and the motion trajectory of the road participant observed by its own sensor is calculated, and the difference is greater than the preset The behavior data of different road participants is used as the first behavior data to achieve the purpose of screening data that is valuable for model training from the behavior data of road participants.

The second type:

The processor 10 determines the behavior category corresponding to the behavior data of each road participant according to the label information corresponding to the behavior data of each road participant, and uses the behavior data of the road participant whose behavior category is a preset category as the first behavior data .

Since there are many types of behaviors of road participants, some of which are more important. If the future trajectory of a road participant predicted by the predictive model is inaccurate, it will further lead to the inaccuracy of the road participant’s behavior. Whether the category is an important behavior category, it may cause a traffic accident, that is, the behavior data of road participants corresponding to the important behavior category is valuable data for model training.

Therefore, when the processor 10 screens the first behavior data, the behavior data of road participants in important behavior categories can be used as the first behavior data, that is, according to the label information corresponding to the behavior data of each road participant, each road participant is determined The behavior category corresponding to the behavior data of the road participant. After the behavior category is determined, the behavior data of the road participant whose behavior category is the preset category is used as the first behavior data. Illustratively, the preset category is lane change behavior or Overtaking behavior.

Therefore, the behavior category corresponding to the behavior data of each road participant is determined according to the label information corresponding to the behavior data of each road participant, and the behavior data of the road participant whose behavior category is the preset category is taken as the first behavior The method of data achieves the purpose of screening data that is valuable for model training from the behavior data of road participants.

The third type:

The processor 10 determines the type of each road participant, and uses the behavior data of the road participant whose type is a preset type as the first behavior data.

Since there are many types of road participants, the trajectory of certain types of road participants may have an impact on the trajectory of other road participants. For example, when driving, most vehicles will be far away from large vehicles. The trajectory of other vehicles may affect the trajectory of other vehicles; or, as pedestrians and two-wheelers are disadvantaged groups, most vehicles will avoid pedestrians and two-wheelers to change the trajectory, thus affecting the trajectory of other road participants The behavioral data of certain types of road participants is valuable data for model training.

Therefore, when the processor 10 filters the first behavior data, it can determine the type of each road participant, and use the behavior data of the road participants whose type is the preset type as the first behavior data. For example, the preset type is Large vehicles, pedestrians or two-wheelers.

Therefore, by judging the type of each road participant, and taking the behavior data of the road participants of the preset type as the first behavior data, it is possible to filter out the behavior data of road participants that is valuable for model training. The purpose of the data.

The cloud 20 receives the first behavior data and corresponding annotation information sent by the processor 10, and stores the received first behavior data and corresponding annotation information in the original database. In order to update the prediction model, a new prediction model needs to be generated. In order to generate a new prediction model, the training samples used by the training model are required. Therefore, the first behavior data can be extracted according to the preset feature extraction method. The amount of feature extraction, the amount of feature extraction and the corresponding label information are stored as training samples in the training sample library. Among them, developers can change the feature extraction method at any time according to their needs.

Since the number of first behavior data and corresponding annotation information sent from the processor 10 each time is limited, it is well known that model training requires a large number of training samples. If only one or several passes are stored in the training sample library Model training with training samples cannot get good training results. Therefore, preset automatic trigger conditions can be set. When the preset automatic trigger conditions are met, the training stored in the training sample library after the last training sample is extracted The sample trains the initial network model to obtain the target network model, where the target network model is used to correlate the feature extraction amount used as the training sample with the corresponding annotation information, and because the feature extraction amount is the representative amount of the behavior data of the road participants The annotation information is the annotation amount of the future motion trajectory. Therefore, the target network model is used to correlate the behavior data of road participants with the corresponding future motion trajectory.

Thus, when the preset automatic trigger conditions are met, the training samples stored in the training sample library after the last training sample was extracted can be used to train the initial network model to obtain the target network model. Therefore, the embodiment of the present invention meets the requirements When the automatic trigger condition is set, the model training can be automatically triggered without manual participation, which greatly reduces labor costs and has a high degree of automation.

Among them, the preset automatic trigger condition can be: the number of training samples stored in the training sample library after the last training sample is extracted reaches the preset number threshold, or the time between the time when the last training sample is extracted and the current time reaches The preset duration.

When the number of training samples stored in the training sample library after the last extraction of the training samples reaches the preset number threshold, it means that the number of training samples has reached the amount of data that can be used for model training. At this time, the cloud 20 extracts the training sample library The training samples stored after the training samples were extracted last time train the initial network model to obtain the target network model.

When the time between the last time the training sample was extracted and the current time reaches the preset time, it means that the number of training samples has grown more and more over time, and it has reached the amount of data that can be used for model training. At this time, the cloud 20 Extract the training samples stored after the last training sample was extracted in the training sample library to train the initial network model to obtain the target network model.

If you want to update the prediction model, you need the new prediction model to have advantages over the prediction model to be updated. Therefore, after the target network model is obtained, the target network model needs to be evaluated according to the preset evaluation method to obtain the evaluation result. When the model update requirement is met, the target network model is sent to the processor 10.

Among them, the cloud 20 evaluates the target network model according to the preset evaluation method to obtain the evaluation result. When the evaluation result meets the model update requirement, the method of sending the target network model to the processor 10 may be:

The cloud 20 predicts the behavior data of each road participant to be predicted in the data set to be predicted based on the target network model to obtain the corresponding future motion trajectory;

According to the future motion trajectory corresponding to the behavior data of each road participant to be predicted and the motion trajectory of each road participant to be predicted observed by its own sensor, the size of the evaluation index of the target network model is calculated;

Calculate the increase amount of the evaluation index of the target network model relative to the evaluation index of the prediction model;

When the amount of increase meets the preset increase requirement, the target network model is sent to the processor 10.

Usually, the performance of the target network model is better than the prediction model to be updated by the increase of the evaluation index. Therefore, it is necessary to calculate the size of the evaluation index of the target network model. The behavior data of each road participant to be predicted is predicted to obtain the corresponding future motion trajectory, and then the future motion trajectory corresponding to the behavior data of each road participant to be predicted and each road participant to be predicted observed by its own sensor Calculate the size of the evaluation index of the target network model.

Exemplarily, the data set to be predicted may be a set of all training samples in the training sample library, or may be a set of other training samples specifically used for evaluation, which is not limited in the embodiment of the present invention. The evaluation index may include the model prediction accuracy rate and/or the absolute error of the model prediction.

After calculating the size of the evaluation index of the target network model, the increase amount of the evaluation index of the target network model relative to the evaluation index of the prediction model can be calculated. When the increase amount meets the preset increase requirement, the performance of the target network model is excellent. For the prediction model to be updated, the target network model can be sent to the processor 10 at this time.

Therefore, after the target network model is obtained, the evaluation is automatically triggered, and the increase in the evaluation index of the target network model relative to the evaluation index of the prediction model meets the preset increase requirement, that is, the performance of the target network model is better than the forecast to be updated When modeling, the target network model is sent to the processor 10 to achieve automatic deployment of the target network model without manual participation, which greatly reduces labor costs and has a high degree of automation.

Since the performance of the target network model is better than the prediction model to be updated, the cloud 20 will send the target network model to the processor 10. Therefore, the processor 10 can update the prediction model as long as it receives the target network model. The prediction model is updated to the target network model. Therefore, after the processor 10 receives the target network model, it can automatically update the prediction model to the target network model, without manual involvement, greatly reducing labor costs, and having a high degree of automation.

It can be seen from the above content that in the embodiment of the present invention, the processor 10 marks the behavior data of each road participant based on the movement trajectory of each road participant observed by its own sensor to obtain the corresponding labeling information, which realizes automatic labeling. Obtain the annotation information instead of manual offline annotation, and then send the filtered first behavior data and the corresponding annotation information to the cloud 20. When the cloud 20 determines that the preset automatic trigger condition is met, it extracts the training sample library from the previous extraction The training samples stored after the training samples train the initial network model to obtain the target network model, which realizes that the model training can be automatically triggered when the preset automatic trigger conditions are met, and then after the target network model is obtained, the evaluation is automatically triggered, and the evaluation is performed When the result meets the model update requirements, the target network model is sent to the processor 10 to achieve the automatic deployment of the target network model. Finally, after the processor 10 receives the target network model, it can automatically update the prediction model to the target network model. It can be seen from the above that no manual participation is required for data labeling, model training, model evaluation, or model update, which greatly reduces labor costs and has a high degree of automation.

In addition, the embodiment of the present invention provides a data-driven system for automatic iteration of predictive models. Only one developer can complete the automatic update of the predictive model, which greatly improves the efficiency of research and development and reduces the cost of research and development. .

Those of ordinary skill in the art can understand that the drawings are only schematic diagrams of an embodiment, and the modules or processes in the drawings are not necessarily necessary for implementing the present invention.

A person of ordinary skill in the art can understand that the modules in the device in the embodiment may be distributed in the device in the embodiment according to the description of the embodiment, or may be located in one or more devices different from this embodiment with corresponding changes. The modules of the above-mentioned embodiments can be combined into one module or further divided into multiple sub-modules.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The technical solutions recorded in the foregoing embodiments are modified, or some of the technical features are equivalently replaced; these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

A data-driven system for realizing automatic iteration of prediction models, characterized in that the system includes a processor and a cloud, the processor is provided with a prediction model, and the prediction model is used to predict the future trajectory of road participants ；

The processor obtains the behavior data of each road participant, where the behavior data includes the environmental static map information at the current moment and the historical movement trajectory of the road participant before the current moment collected by the collection device installed in itself, based on The motion trajectory of each road participant observed by its own sensor is labeled with the behavior data of each road participant to obtain the corresponding label information, and the behavior data of each road participant is selected from the behavior data of each road participant that meets the preset screening requirements First behavior data, sending the first behavior data and corresponding annotation information to the cloud;

The cloud stores the received first behavior data and corresponding annotation information in the original database, and performs feature extraction on the first behavior data according to a preset feature extraction method to obtain a feature extraction amount, and the feature extraction amount is And the corresponding labeling information are stored as training samples in the training sample library, and when the preset automatic trigger conditions are met, the training samples stored in the training sample library after the last training sample is extracted are extracted to train the initial network model to obtain the target A network model, wherein the target network model is used to correlate the behavior data of road participants with the corresponding future motion trajectory, and evaluate the target network model according to a preset evaluation method to obtain the evaluation result, when the evaluation result When the model update requirement is met, sending the target network model to the processor;

The processor receives the target network model, and updates the prediction model to the target network model.
The system according to claim 1, wherein the processor uses the movement trajectory of each road participant observed by its own sensor as the label information corresponding to the behavior data of each road participant.
The system of claim 1, wherein the processor predicts and obtains the future motion trajectory corresponding to the behavior data of each road participant based on the prediction model, and calculates the road participant for each road participant The difference between the future motion trajectory corresponding to the behavior data of and the motion trajectory of the road participant observed by its own sensor, and the behavior data of the road participant whose difference is greater than the preset difference is used as the first behavior data;

or,

The processor determines the behavior category corresponding to the behavior data of each road participant according to the label information corresponding to the behavior data of each road participant, and uses the behavior data of the road participants whose behavior category is a preset category as the first One behavioral data;

or,

The processor determines the type of each road participant, and uses the behavior data of the road participant whose type is a preset type as the first behavior data.
The system according to claim 3, wherein the preset category is lane changing behavior or overtaking behavior.
The system of claim 3, wherein the preset type is a large vehicle, a pedestrian, or a two-wheeled vehicle.
The system according to claim 1, wherein the processor stores the first behavior data and corresponding annotation information before sending the first behavior data and corresponding annotation information to the cloud.
The system according to claim 1, wherein when the number of training samples stored in the training sample library after the last training sample is extracted reaches a preset number threshold, the cloud extracts the training sample from the training sample library. Training the initial network model with the training samples stored after the training samples were extracted last time to obtain the target network model;

or,

When the time between the last time the training sample was extracted and the current time reaches the preset time length, the cloud extracts the training samples stored in the training sample library after the last time the training samples were extracted to train the initial network model to obtain the target Network model.
The system according to claim 1, wherein the cloud predicts the behavior data of each road participant to be predicted in the data set to be predicted based on the target network model to obtain the corresponding future motion trajectory;

Calculate the size of the evaluation index of the target network model according to the future motion trajectory corresponding to the behavior data of each road participant to be predicted and the motion trajectory of each road participant to be predicted observed by its own sensor;

Calculating an increase amount of the evaluation index of the target network model relative to the evaluation index of the prediction model;

When the amount of increase meets a preset increase requirement, the target network model is sent to the processor.
The system of claim 1, wherein the road participants include vehicles and/or pedestrians.