A training method and system for environmental vehicle behavior prediction model
By using on-board sensors and cloud databases to screen and annotate data in autonomous vehicles, an automated model training process is built, which solves the problems of low automation and low training efficiency in the existing technology, and realizes efficient automated training of predictive models.
Patent Information
- Application Number
- CN202210454730.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-24
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-04-24
AI Technical Summary
In the prior art, there are problems such as low degree of automation, high manual intervention and low training efficiency of automated training prediction models.
Vehicle data is obtained through on-board sensors, processed and uploaded to the cloud, and data screening and annotation are used for databases and feature engineering, and an automated model training process is built to improve the degree of automation of the training of prediction models, reduce manual intervention, and improve training efficiency.
It realizes efficient automated training of predictive models, reduces manual intervention, improves training efficiency, and improves the degree of automation of model training.
Smart Images

Figure CN114742236B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of autonomous driving technology, and in particular to a method and system for training an environmental vehicle behavior prediction model. Background Art
[0002] The prediction of environmental target vehicle behavior is of great significance for the safe and efficient control of autonomous driving vehicles. Compared with the prediction of manually adjusted parameter models that require expert knowledge and fixed scene adaptation, the artificial intelligence machine learning method training model has the generalization and robustness to adapt to multiple scenes. Artificial intelligence methods are gradually favored by autonomous driving engineering implementation, among which the prediction model training process is an important link, including the recording and collection of driving data, transmission and storage, sorting and cleaning, analysis and annotation, and driving training. This method is designed and implemented accordingly to solve the problem of automatic and efficient prediction model training.
[0003] Patent CN2020101227910 provides a prediction model generation method for an intelligent network cloud platform based on reinforcement learning. The patent uses an intelligent network car cloud platform to record and upload user operation records and car operation status. In addition to driving data, operator selection is also included in the training process. Although this method expands the scope of model training, it cannot effectively improve the efficiency of model training by only increasing computing power to expand the training scale and automatically screen operators to update the model structure. Expert knowledge is still important for operator selection and arrangement in model design, and this method does not solve the problem of efficient processing of training data screening and annotation. Summary of the invention
[0004] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a method and system for training an environmental vehicle behavior prediction model to solve the problems of low automation, high manual intervention and low training efficiency in the process of automated training of prediction models in the prior art.
[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0006] A method for training an environmental vehicle behavior prediction model comprises the following steps:
[0007] S1: Vehicle data information is obtained through various vehicle-mounted sensors, and the vehicle-mounted terminal processes the vehicle data information to obtain vehicle driving data;
[0008] S2: The vehicle terminal uploads the driving data to the cloud;
[0009] S3: The cloud receives the driving data uploaded by each vehicle terminal and enters the driving data into the database for persistence;
[0010] S4: The training server extracts driving data from the database in the cloud and processes the driving data to obtain training data;
[0011] S5: Label the training data according to the model structure to be trained;
[0012] S6: Add the training data annotated in S5 to the training set, and train the prediction model to be trained, so that the prediction model training achieves the expected effect;
[0013] S7: Send the prediction model trained in S6 to the test vehicle.
[0014] The present invention also provides an environmental vehicle behavior prediction model training system, characterized in that it includes a collection module, a vehicle-side edge computer system, a vehicle-cloud wireless service bus, a database module, a data screening module, a data annotation module and a model training module; wherein the collection module is used to collect data information obtained by various sensors of the vehicle and send the data information to the vehicle-side edge computer system;
[0015] The vehicle-side edge computer system is used to calculate the vehicle's driving data based on the data sent by the acquisition module, and select the data required by the prediction model to be trained according to the structure of the prediction model to be trained, and upload it to the database module in the cloud through the vehicle-cloud wireless service bus;
[0016] The database module is used to upload data through the car-cloud wireless service bus, encapsulate the received data and enter it into the database, and send the required driving data to the data screening module according to the request sent by the training server;
[0017] The data screening module is used to receive the driving data sent by the database module, screen the received driving data, delete the driving data that does not meet the characteristics required by the prediction model input data, and then send the driving data to the data annotation module;
[0018] The data annotation module is used to receive the driving data sent by the data screening module, automatically annotate the driving data, and send the annotated data to the model training module;
[0019] The model training module is used to receive the data sent by the data annotation module, and continuously train the trained prediction model until the prediction model achieves the expected effect, and then sends it to the test vehicle through the vehicle-cloud service bus.
[0020] Compared with the prior art, the present invention has the following beneficial effects:
[0021] The method and system described in the present invention comprehensively use methods such as vehicle networking, database, feature engineering, and machine learning to build an automated model training process from the collection and transmission of driving data to the cleaning, annotation, and use of training data, thereby improving the degree of automation in training the prediction model, reducing manual intervention, and improving training efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 It is a logic flow chart of the present invention. DETAILED DESCRIPTION
[0023] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present invention. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.
[0024] 1. A training method for environmental vehicle behavior prediction model
[0025] S1: Vehicle data information is obtained through various on-board sensors, and the on-board terminal processes the vehicle data information to obtain the vehicle's driving data. The on-board terminal packages each frame of driving data with a timestamp to form a time series of driving data. Among them, the vehicle side is mainly a test vehicle or a user vehicle. The various sensors deployed on the operating vehicle include but are not limited to visible light cameras, infrared cameras, lidars, millimeter-wave radars, acoustic radars, multi-mode positioning, etc. The data of multiple sensors are inferred and identified by the vehicle-side edge computing system, and finally the vehicle's positioning, speed, heading and other vehicle information are obtained, and the position, heading, speed and other information of the surrounding target vehicles after sensor fusion are obtained. Each frame of information is timestamped and packaged to form time series data.
[0026] S2: The vehicle terminal uploads the driving data to the cloud. The vehicle terminal selects the driving data according to the data required by the prediction model to be trained and uploads it to the cloud. In addition, the vehicle terminal inputs the unique vehicle number corresponding to the driving data and encrypts it. The driving data used for model training is initially packaged and uploaded to the cloud server through the vehicle-side network cloud. The unique vehicle number is packaged and encrypted during the upload process to ensure both security and efficiency.
[0027] S3: The cloud receives the driving data uploaded by each vehicle terminal and enters the driving data into the database for persistence. The cloud server receives the data uploaded by the vehicle terminal in each vehicle. The original driving data of the cloud server enters the database for persistence.
[0028] S4: The training server extracts driving data from the cloud database and processes the driving data to obtain training data. In the physical process, the training server reorganizes the persistent driving data to obtain the data structure and file organization format required for training. At the same time, the driving data is screened according to the characteristics required by the prediction model input data, and the driving data that does not meet the characteristics is deleted.
[0029] S5: Label the training data according to the model structure to be trained. The training data is structured as a time series, and the data predicted for the future moment is labeled with the historical data of the next step time window. A sampling data sequence is organized to automatically label the training data.
[0030] S6: Add the training data annotated in S5 to the training set, and train the prediction model to be trained, so that the prediction model training achieves the expected effect;
[0031] 2. A training system for environmental vehicle behavior prediction models
[0032] The system includes a collection module, a vehicle-side edge computer system, a vehicle-cloud wireless service bus, a database module, a data screening module, a data annotation module and a model training module; wherein the collection module is used to collect data information obtained by various sensors of the vehicle and send the data information to the vehicle-side edge computer system;
[0033] The vehicle-side edge computer system is used to calculate the vehicle's driving data based on the data sent by the acquisition module, and select the data required by the prediction model to be trained according to the structure of the prediction model to be trained, and upload it to the database module in the cloud through the vehicle-cloud wireless service bus;
[0034] The database module is used to upload data through the car-cloud wireless service bus, encapsulate the received data and enter it into the database, and send the required driving data to the data screening module according to the request sent by the training server;
[0035] The data screening module is used to receive the driving data sent by the database module, screen the received driving data, delete the driving data that does not meet the characteristics required by the prediction model input data, and then send the driving data to the data annotation module;
[0036] The data annotation module is used to receive the driving data sent by the data screening module, automatically annotate the driving data, and send the annotated data to the model training module. The data annotation module annotates according to the different data required to be input by the prediction model structure to be trained.
[0037] The model training module is used to receive data sent by the data annotation module, and continuously train the trained prediction model until the prediction model achieves the expected effect, and then sends it to the test vehicle through the vehicle-cloud service bus. In the process of training the prediction model, the model training module can simultaneously receive new annotated driving data and automatically add it to the training set for training; the weight of the prediction model during the training process can be saved in the model training module in real time.
[0038] Combined with Figure 1 , the method and system of the present invention are explained:
[0039] In the overall process, the test vehicle data acquisition, collection, organization, storage, data labeling, model training, and model updating are completed in a closed-loop automatic process.
[0040] 1 and 2 are both part of the vehicle-side edge computer system. 1 is the sensor acquisition module of the vehicle. The collected data is used in 2 to perform edge computing, calculate the sensor data, and reconstruct the vehicle environment information. After the calculation here, the data required for behavior prediction model training is obtained, such as the location trajectory of surrounding vehicles, driving behavior, and the trajectory and speed of the vehicle. This information is related to the network structure design of the prediction model and is selectively sent upstream.
[0041] 3 is the network path for uplink and downlink data from the vehicle to the cloud. It is based on the wireless network layer of the Internet of Vehicles, and the underlying network guarantees bandwidth and security.
[0042] 4. The vehicle-cloud wireless service bus is the vehicle-cloud networking server system, which completes the network layer server functions and provides bus-shaped functions to upstream and downstream applications. The upstream accesses the downstream vehicle data through the bus, and the downstream vehicle publishes data to the upstream. The cloud service bus encapsulates the underlying network process and provides a reliable and secure interface.
[0043] 5 is a database module connected to the car-cloud bus. This module encapsulates the functions of collecting, sorting and storing upstream and downstream data. After the downstream car-side data is uploaded via the cloud service bus, the original data is stored in the database module for subsequent data annotation processing. After the unqualified data is eliminated by the 6 data screening module, it can be stored in the database.
[0044] 6. Data screening executes the process of eliminating unqualified data and performs feature engineering processing based on the input data requirements of the prediction model.
[0045] 7. Automatically label the data and process them according to the different prediction model structures. For example, use the dynamic model to generate the predicted position information of several moments in the future as training data based on the trajectory information of several moments in the past, or label the first half of the position information of a time series as historical trajectory information and the second half as predicted trajectory information based on the actual trajectory information.
[0046] 8 is a model training module. The module input information is the training data that has been annotated in 7. When new annotated data arrives during the training process, it is automatically added to the training set for continuous training.
[0047] 9. When the model training reaches a certain expected effect, the updated model is pushed to the running test vehicle through the vehicle-cloud service bus.
[0048] 10: The model network weights can be saved during the training process.
[0049] As described above, the reminder system of the present invention is not limited to the configuration described above, and other systems that can implement the embodiments of the present invention may fall within the scope of protection of the present invention.
[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.
Claims
1. A training method for an environmental vehicle behavior prediction model, It is characterized in that The steps include: S1: Vehicle data information is obtained through various vehicle-mounted sensors, and the vehicle-mounted terminal processes the vehicle data information to obtain vehicle driving data; S2: The vehicle terminal uploads the driving data to the cloud; S3: The cloud receives the driving data uploaded by each vehicle terminal and enters the driving data into the database for persistence; S4: The training server extracts driving data from the database in the cloud and processes the driving data to obtain training data; S5: Label the training data according to the model structure to be trained; S6: Add the training data annotated in S5 to the training set, and train the prediction model to be trained, so that the prediction model training achieves the expected effect; S7: Send the prediction model trained in S6 to the test vehicle; The method is used to automatically train the prediction model to be trained, and send the prediction model obtained after training to the test vehicle; The method is implemented through an environmental vehicle behavior prediction model training system, which includes an acquisition module, a vehicle-side edge computer system, a vehicle-cloud wireless service bus, a database module, a data screening module, a data annotation module and a model training module; wherein the acquisition module is used to collect data information obtained by various sensors of the vehicle and send the data information to the vehicle-side edge computer system; The vehicle-side edge computer system is used to calculate the vehicle's driving data based on the data sent by the acquisition module, and select the data required by the prediction model to be trained according to the structure of the prediction model to be trained, and upload it to the database module in the cloud through the vehicle-cloud wireless service bus; The database module is used to upload data through the car-cloud wireless service bus, encapsulate the received data and enter it into the database, and send the required driving data to the data screening module according to the request sent by the training server; The data screening module is used to receive the driving data sent by the database module, screen the received driving data, delete the driving data that does not meet the characteristics required by the prediction model input data, and then send the driving data to the data annotation module; The data annotation module is used to receive the driving data sent by the data screening module, automatically annotate the driving data, and send the annotated data to the model training module; The model training module is used to receive the data sent by the data annotation module, and continuously train the trained prediction model until the prediction model achieves the expected effect, and then sends it to the test vehicle through the vehicle-cloud service bus.
2. According to the environmental vehicle behavior prediction model training method of claim 1, It is characterized in that In S1, the vehicle terminal packages the timestamp of each frame of driving data to form a time series of driving data.
3. According to the environmental vehicle behavior prediction model training method of claim 1, It is characterized in that In S2, the vehicle terminal inputs the unique vehicle number corresponding to the driving data and encrypts it.
4. The environmental vehicle behavior prediction model training method according to claim 1, It is characterized in that In S2, the vehicle terminal filters the driving data according to the data required by the prediction model to be trained and uploads it to the cloud.
5. The environmental vehicle behavior prediction model training method according to claim 1, It is characterized in that In S4, the persistent driving data is reorganized to obtain the data structure and file organization format required for training.
6. The environmental vehicle behavior prediction model training method according to claim 1, It is characterized in that In S4, the driving data is screened according to the characteristics required by the prediction model input data, and the driving data that does not meet the characteristics is deleted.
7. The environmental vehicle behavior prediction model training method according to claim 1, It is characterized in that In S5, the training data is structured as a time series, and the data for predicting future moments is annotated with the historical data of the next step time window. A sampling data sequence is organized to automatically annotate the training data.
8. The environmental vehicle behavior prediction model training method according to claim 1, It is characterized in that The data annotation module annotates the data according to the different input data required by the prediction model structure to be trained.
9. The environmental vehicle behavior prediction model training method according to claim 8, It is characterized in that During the training process of the prediction model, the model training module can simultaneously receive new labeled driving data and automatically add it to the training set for training; the weight of the prediction model during the training process can be saved in the model training module in real time.
