Driving model training method and device and storage medium

By training the driving model, integrating the traffic environment and driving behavior, the problem of insufficient accuracy in driving style recognition in the assisted driving system is solved, personalized driving experience and environmental adaptability are improved, and user experience is optimized.

CN118981633APending Publication Date: 2024-11-19ZHEJIANG HUARUIJIE TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410843518.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-26
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

At this stage, the driving style recognition in the assisted driving system is insufficient, and the personalized configuration is not effective, making it difficult to adapt to the driving habits of different drivers.

Method used

By determining the correspondence between traffic participant information and driving behavior parameters around the vehicle, training driving models, integrating traffic environment information and driving behavior, forming a fine driving behavior database, enhancing situational perception capabilities, and providing personalized driving experience.

Benefits of technology

It improves the environmental adaptability of the assisted driving system and the personalized matching of users, reduces the frequency of manual takeovers, and optimizes the user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981633A_ABST
    Figure CN118981633A_ABST
Patent Text Reader

Abstract

The invention discloses a driving model training method and device and a storage medium, and the method comprises the steps: determining the driving behavior data of a vehicle, and enabling the driving behavior data to comprise the corresponding relation between the information of traffic participants around the vehicle and the driving behavior parameters of the vehicle, the traffic participant information comprises types and speeds of traffic participants; and training a driving model based on the driving behavior data of the vehicle. Through the mode, the environment adaptability and the user personalized matching degree of the auxiliary driving system can be improved, so that the manual taking-over frequency is reduced, and the user experience is optimized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to a driving model training method, device and storage medium. Background Art

[0002] Intelligence is one of the important development trends in the automotive industry today. With the rapid progress of artificial intelligence, big data, cloud computing, sensor technology and wireless communication technology, intelligent driving systems have evolved from preliminary assisted driving functions, such as blind spot monitoring and adaptive cruise control, to a stage that can achieve partial or even highly automated driving.

[0003] At present, the research and development of autonomous driving technology and systems are mostly based on two application scenarios: highways and parking. The system composition usually includes sub-modules such as perception, positioning, planning, and decision-making control. For structured highway and urban scene applications, intelligent driving systems are evolving from lane-level cruising and congestion assistance to high-level point-to-point autonomous driving in specific scenarios. Human-machine co-driving will be an important application scenario that intelligent driving systems will face for a long time in the future. Therefore, driving style learning and configuration have become an indispensable part of intelligent driving systems. However, the accuracy of driving style recognition and the effect of achieving personalized configuration in assisted driving systems at this stage need to be improved. Summary of the invention

[0004] In order to solve the above technical problems, the technical solution adopted in this application is: to provide a driving model training method, device and storage medium to at least solve the problems of insufficient accuracy in driving style recognition and poor personalized configuration effect.

[0005] According to one embodiment of the present invention, a driving model training method is provided, comprising:

[0006] Determining driving behavior data of the vehicle, the driving behavior data including a correspondence between traffic participant information around the vehicle and a driving behavior parameter of the vehicle, the traffic participant information including a type and a speed of the traffic participant;

[0007] A driving model is trained based on the driving behavior data of the vehicle.

[0008] In order to solve the above technical problems, a technical solution adopted in the present application is: to provide a device, including a memory and a processor, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, it is used to implement the driving model training method in the above technical solution.

[0009] In order to solve the above technical problems, a technical solution adopted in the present application is: providing a computer-readable storage medium, which is used to store a computer program. When the computer program is executed by a processor, it is used to implement the driving model training method in the above technical solution.

[0010] Through the above scheme, the beneficial effect of the present application is: the driving model training method provided by the present application obtains the driving behavior data of the vehicle by determining the correspondence between the traffic participant information around the vehicle (including category and speed) and the driving behavior parameters of the vehicle, and trains the driving model based on the driving behavior data of the vehicle; thereby, it is possible to integrate the traffic environment information around the vehicle and the driving behavior of the vehicle itself to form a comprehensive and detailed driving behavior database; the driving model trained based on the database can enhance situational awareness, identify and adapt to the driving habits of different drivers to provide a more personalized driving experience, further improve the environmental adaptability and user personalization matching of the assisted driving system, reduce the frequency of manual takeover, and optimize the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0012] Figure 1 It is a hardware structure diagram of an intelligent transportation system provided by this application;

[0013] Figure 2 It is a flowchart of an embodiment of a driving model training method provided by the present application;

[0014] Figure 3 is a schematic diagram of a video analysis network provided by the present application;

[0015] Figure 4 is a schematic diagram of a state transfer matrix provided by the present application;

[0016] Figure 5 is a schematic diagram of an observation matrix provided by this application;

[0017] Figure 6 is a flowchart of another embodiment of the driving model training method provided by the present application;

[0018] Figure 7 It is a structural schematic diagram of an embodiment of a device provided by the present application;

[0019] Figure 8It is a structural schematic diagram of an embodiment of a computer-readable storage medium provided by the present application. DETAILED DESCRIPTION

[0020] The present application is further described in detail below in conjunction with the accompanying drawings and examples. It is particularly noted that the following examples are only used to illustrate the present application, but are not intended to limit the scope of the present application. Similarly, the following examples are only some embodiments of the present application rather than all embodiments, and all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0021] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0022] It should be noted that, in the description of the present application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically limited. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0023] This application provides an intelligent transportation system using the driving model training method. Figure 1 , Figure 1 This is a hardware structure diagram of an intelligent transportation system provided in this application, and the intelligent transportation system includes an on-board edge device and a cloud server.

[0024] The cloud server communicates with the edge controller in the vehicle-mounted edge device through wireless communication technology or other feasible methods, including signal acquisition unit, data storage unit, AI training unit, redundant power supply and / or vehicle-cloud communication unit, etc. Among them, the AI ​​training unit can use machine learning and deep learning algorithms to process and analyze the collected data, which can be used to train driving models.

[0025] The vehicle-mounted edge device includes a perception unit, an edge controller and / or an execution / interaction unit, etc., specifically including:

[0026] (1) The perception unit is responsible for collecting information about the surrounding environment and the internal state of the system, including: a. visual sensors such as forward-looking cameras, which are used to capture road images in the direction of the vehicle's travel and identify lane lines, traffic signs, pedestrians, other vehicles, etc.; and / or, b. sensors that provide vehicle dynamics signals, wherein the vehicle dynamics signals include vehicle speed, acceleration, steering angle, throttle and brake status, etc., to help the system understand the vehicle's own motion state; and / or, c. millimeter-wave radar sensors, which use electromagnetic waves in the millimeter-wave frequency band to detect surrounding objects, can penetrate rain, fog, and dust, and provide reliable distance, speed, and angle information, which is particularly suitable for obstacle detection in severe weather conditions.

[0027] The components in the perception unit work together to build a comprehensive view of the vehicle's surroundings by fusing the data they collect, supporting the system to make timely and accurate driving decisions. In addition, with the development of technology, some systems can also integrate lidar sensors, ultrasonic sensors, infrared sensors, etc. to further enhance perception capabilities and ensure the safety and reliability of intelligent transportation systems.

[0028] (2) Edge controller, which may also be referred to as Intelligent Driving Domain Controller (IDDC) in the intelligent transportation system proposed in this application, is a core component in the vehicle's electrical and electronic architecture (E / E Architecture), and is designed specifically to realize advanced driver assistance systems (ADAS) and autonomous driving (AD) functions. It specifically includes a signal acquisition unit, an AI reasoning unit, an information fusion unit, a vehicle-cloud communication unit, a microcontroller MCU, Flash / eMMC, a power supply and / or memory, etc.

[0029] The vehicle's intelligent driving domain controller is a high-performance computing platform that integrates a powerful processor. It collects visual and / or radar sensor signals through high-speed transmission technologies such as low-voltage differential signal LVDS (Low-Voltage Differential Signaling) and controller area network bus (CAN, Controller Area Network), processes data from various sensors in real time, and executes complex algorithms and AI models to achieve rapid generation of decisions and control instructions.

[0030] (3) Execution / interaction unit, which is responsible for receiving instructions from the edge controller via CAN and / or Ethernet, performing corresponding operations based on these instructions, and collecting environmental or system status data and feeding it back to the edge controller. Specifically, it includes the human-machine interface (HMI) and actuators, such as steering, braking, throttle and other control systems, so that the driver (or direct operator in the case of unmanned driving) can receive system feedback and / or intervene in control.

[0031] This application provides a driving model training method based on the above-mentioned intelligent transportation system, which obtains the driving behavior data of the vehicle by collecting the corresponding relationship between the information of traffic participants around the vehicle and the driving behavior parameters of the vehicle, and trains the driving model based on the driving behavior data of the vehicle. In this way, the traffic environment information around the vehicle and the driving behavior of the vehicle itself are integrated to form a comprehensive and detailed driving behavior database. The driving model trained based on this database can enhance the situational awareness ability, identify and adapt to the driving habits of different drivers to provide a more personalized driving experience, further improve the environmental adaptability and user personalized matching of the assisted driving system, reduce the frequency of manual takeover, and optimize the user experience.

[0032] See also Figure 2 , Figure 2 is a flow chart of an embodiment of the driving model training method provided by the present application. It should be noted that if there are substantially the same results, this embodiment does not Figure 2 The process sequence shown is limited. Figure 2 As shown, this embodiment includes:

[0033] S110: Determine driving behavior data of the vehicle, wherein the driving behavior data includes a correspondence between traffic participant information around the vehicle and driving behavior parameters of the vehicle, wherein the traffic participant information includes a type and speed of the traffic participant.

[0034] Through the perception unit integrated in the intelligent vehicle or traffic management system, for example, by using visual sensors (such as cameras) and millimeter wave radar sensors to work together, comprehensive and real-time monitoring of the traffic environment around the vehicle is achieved to perceive the information of traffic participants around the vehicle. Exemplarily, the edge (intelligent driving domain) controller receives and pre-processes visual and / or radar sensor signals through LVDS and / or CAN buses to determine information such as the category, speed and / or distance of traffic participants.

[0035] In one embodiment, the visual sensors installed on the vehicle perceive the video images of traffic participants in the 360° scene around the vehicle. After pre-processing these video signals, the intelligent driving domain controller uses a deep learning model such as a multi-task convolutional neural network (CNN) to simultaneously process multiple tasks in the image, including object recognition, position estimation, and speed prediction, and outputs information such as the category, speed, and distance of the traffic participants.

[0036] In one example, the edge controller receives video signals, i.e., raw image data, from cameras and / or other visual sensors at high speed through an LVDS interface. These data contain rich scene information, such as road conditions, pedestrians, other vehicles, etc. The received video signals are preprocessed so that the controller can use image processing algorithms to perform preliminary analysis on the raw image data, such as image enhancement, feature extraction, target detection, etc., to quickly identify and track the category, speed, and position of traffic participants. The preprocessing of video signals includes image scaling, ROI (region of interest) cropping, channel cropping and format conversion, and brightness normalization.

[0037] The video data around the vehicle is analyzed using a video analysis network to obtain the type, speed and distance of traffic participants in the video data; and / or, the radar signal collected by the vehicle is processed to obtain the distance and speed of the traffic participants.

[0038] In one embodiment, a semantic, geometric and motion joint analysis is performed on traffic participants in a 360° scene around the vehicle based on a video analysis network (multi-task convolutional neural network), and after post-processing such as maximum suppression, threshold filtering and bandpass filtering, the category, distance and speed information of the traffic participants are output.

[0039] For example, see Figure 3 , Figure 3 It is a schematic diagram of a video analysis network provided by the present application. The video analysis neural network may include multiple convolutional layers, pooling layers, fully connected layers and / or output layers, and can perform multiple related but different tasks simultaneously.

[0040] Integrate road scene images from multiple directions (such as front, rear, side view cameras and / or radar data fusion), use the first stage of the convolutional neural network to extract common features (i.e. shared features) of all input data, and further perform semantic, geometric and motion joint analysis through hierarchical processing of high-level understanding analysis, such as identifying which type of object each pixel or region belongs to (such as road, vehicle or pedestrian, etc.), determining the specific position (x, y) of each target through bounding box prediction or instance segmentation, and estimating the dynamic attributes of each traffic participant such as speed (u, v). After the analysis results are subjected to post-processing processes such as maximum suppression, threshold filtering and bandpass filtering, the categories of traffic participants (1-car, 2-bus, 3-truck, 4-two-wheeled vehicle, 5-three-wheeled vehicle, 6-special vehicle, 7-other vehicles), distance (x, y) and speed information (u, v) are output.

[0041] In another example, the edge controller communicates with radar sensors and / or other vehicle sensors (such as ultrasonic sensors, LiDAR, etc.) through the CAN bus to collect more dimensional information about the surrounding environment. After preprocessing the CAN signal, the controller can integrate these multi-source sensor information and improve the accuracy and robustness of environmental perception through data fusion technology. Among them, the preprocessing of CAN signals includes signal ID filtering, signal parsing, and input structure reconstruction.

[0042] Optionally, based on the multi-transmit and multi-receive 77GHz millimeter-wave radar signal input, after Fourier transform, target clustering, ranging, angle measurement and tracking, the distance (x, y) and speed (u, v) information of the traffic participants are output.

[0043] In addition, the present application may also select or simultaneously use other sensing devices that can be used to sense information of traffic participants around the vehicle to obtain information such as the type, speed and distance of traffic participants, which will not be repeated here.

[0044] In one embodiment, at least two perception methods are used to detect the speed of traffic participants to obtain at least two perceived speeds of traffic participants; and / or, at least two perception methods are used to detect the distance of traffic participants to obtain at least two perceived distances of traffic participants.

[0045] In one example, based on the Hungarian matching algorithm and the kinematic Kalman filter fusion, at least two perceived speeds of the traffic participant are fused to obtain the speed of the traffic participant; based on the Hungarian matching algorithm and the kinematic Kalman filter fusion, at least two perceived distances of the traffic participant are fused to obtain the distance of the traffic participant.

[0046] Optionally, considering the different time differences between the multi-sensor inputs, the time perception result output by the video analysis network is fused with the millimeter wave radar perception result. Exemplarily, based on the perception time difference between at least two perception methods, at least two perceived speeds of the traffic participant are fused to obtain the speed of the traffic participant; and / or, based on the perception time difference between at least two perception methods, at least two perceived distances of the traffic participant are fused to obtain the distance of the traffic participant.

[0047] Before multi-sensor data fusion, the Hungarian matching algorithm is first used to solve the correspondence between traffic participants observed by multiple sensors at the same time. Based on the principle of cost minimization, it ensures that each visually perceived traffic participant is optimally matched with a radar-perceived traffic participant, or in a multi-traffic participant scenario, the observation data of each sensor is correctly associated. Kalman filtering is a recursive estimation method suitable for state estimation of dynamic systems with noise. It is used to fuse visual and radar data to obtain more accurate distances (x, y) and speeds (u, v) of traffic participants.

[0048] In one example, the specific process is as follows:

[0049] 1. Initialization: At time step k = 0, initialize the state vector X 0|0 and the state estimation covariance matrix P 0|0 . Among them, X=[x, y, u, v] T , where x and y are the position coordinates, and u and v are the velocity components.

[0050] 2. Prediction step (for each time step k):

[0051] a. Use formula (1) to use the state transfer matrix F k To predict the state vector for the next time step:

[0052] X k|k-1 =F k X k-1|k-1 (1)

[0053] Among them, X k-1|k-1 is the state estimate for time step k-1, which means the best estimate based on the available information at the end of time step k-1; F k is the state transfer matrix for time step k, exemplarily, as Figure 4 As shown, F k Describes how the state evolves in time, including dynamics model and time delay compensation, where t is the Kalman iteration period, tc is the visual sensor time compensation, and tl is the radar sensor time compensation; X k|k-1 is the predicted state at time step k, based on F kand X k-1|k-1 It is calculated that the observation data at time step k is not considered at this time.

[0054] b. Use formula (2) to update the predicted state estimation covariance matrix:

[0055] P k|k-1 =F k P k-1|k-1 F k T + Q k (2)

[0056] Among them, Q k is the process noise matrix at time step k, representing the uncertainty or internal noise in the state transition process; P k|k-1 is the estimated covariance matrix of the predicted state at time step k, representing the predicted state X k|k-1 uncertainty.

[0057] 3. Update step (for each time step k):

[0058] a. Calculate the Kalman gain using formula (3):

[0059] K k =P k|k-1 H k T (H k P k|k-1 H k T + R k ) -1 (3)

[0060] b. Collect observation data Z at time step k k ;

[0061] c. Calculate the prediction error Z k -H k X k|k-1 , reflecting the inconsistency between predictions and actual observations;

[0062] d. Update the state vector using the Kalman gain and prediction error using formula (4):

[0063] X k|k =X k|k-1 + K k (Z k -H k X k|k-1 ) (4)

[0064] e. Update the state estimation covariance matrix using formula (5) to reflect the impact of new information:

[0065] Pk|k =P k|k-1 (I - K k H k ) (5)

[0066] In the above update steps, H k is the observation matrix of time step k, such as Figure 5 As shown, it is used to map the state vector to the observation space and define the relationship between the observation and the state; Z k is the observation vector at time step k, which contains all available observation data, which may come from different sensors; R k is the observation noise matrix at time step k, representing the uncertainty or noise in the observation data; K k is the Kalman gain at time step k, which is a weight vector used to determine the relative importance of observation data in state estimation update; X k|k is the updated state at time step k, which is the fusion prediction state X k|k-1 and observation data Z k The result after that; k|k is the updated state estimate covariance matrix for time step k, representing the updated state X k|k uncertainty; I is the unit matrix, which is used in the covariance update equation and indicates that the diagonal elements of the covariance matrix will not be changed.

[0067] 4. Iteration: Repeat steps 2 and 3 until all time steps are iterated.

[0068] 5. Output: At each time step k, output the updated state vector X k|k and the state estimation covariance matrix P k|k .

[0069] Through the above iterative process, the Kalman filter can effectively fuse the data from multiple sensors and provide continuous and accurate state estimation. In addition, in addition to the above fusion methods, weighted fusion or artificial neural network methods can also be selected to fuse the perception results of multiple sensors to obtain more accurate information such as the distance and speed of traffic participants, which is not limited here.

[0070] Based on the above fused perception results, more accurate distances (x, y) and speeds (u, v) of traffic participants are obtained, so as to more accurately calculate the driving behavior parameters of the vehicle, such as the longitudinal following distance and lateral distance target distance of the vehicle, etc. The driving behavior parameters of the vehicle are quantitative indicators that describe the driver's operating habits, vehicle dynamic response and driving safety during vehicle driving, including longitudinal following distance statistics, acceleration and deceleration statistics, lateral distance target distance statistics, lane change yaw rate statistics and yaw rate change rate statistics, etc. These parameters can reflect the driver's operating style, vehicle performance and road driving safety.

[0071] Based on the relevant sensors configured on the vehicle, such as accelerometers, gyroscopes and angular velocity sensors, distance sensors, wheel speed sensors or inertial measurement units (IMUs), they are sent to the edge controller through the CAN bus or other vehicle networks to obtain the vehicle's dynamic signals such as vehicle speed, acceleration, steering angle, yaw angular velocity, etc.

[0072] In one embodiment, different vehicle behavior parameters are calculated based on different identified driving scenarios, and the vehicle driving scenarios are mainly divided into following scenarios and lane changing scenarios. In one example, the current vehicle driving scenario is determined based on whether the left and right lane line IDs have changed, for example: a. If the left and right lane line IDs have not changed, and the IDs of the traffic participants in front have not changed, it is determined to be a following scenario; b. If the left and right lane line IDs have changed, and the lateral distance of the lane line on the changed side changes continuously and the sign switches, it is determined to be a lane changing scenario.

[0073] In the following vehicle scenario, obtain the longitudinal following vehicle distance statistics hwd i,j , Acceleration and deceleration statistics a i,j , lateral vehicle distance keeping target distance statistics value y i,j , which can be calculated by the following formula:

[0074]

[0075]

[0076] Where i in formula (6) is the category of the traffic participant in the vehicle's forward direction, i∈{1,2,...,7}, j is the speed of the traffic participant j∈{30,60,...,120}, hwd i,j,t is the longitudinal following distance at a certain time t, calculated as hwd i,j,t=x / v_ego, where x is represented by the distance x of the traffic participant in the fusion perception result, and v_ego is the speed of the vehicle at time t; the meanings of i and j in formula (7) are the same as those in formula (1), and a i,j,t is the acceleration and deceleration value of the vehicle at time t, which can be obtained from the vehicle CAN bus; i in formula (8) is the category of the lateral traffic participant i∈{1,2,...,7}, the speed of the lateral traffic participant j∈{30,60,...,120}, y i,j,t It is represented by the lateral distance y of the traffic participant in the aforementioned fusion perception result.

[0077] In the lane change scenario, the lane change yaw rate statistic ω in the driving behavior parameter is calculated by the following formula (9): i And the yaw rate change statistics Where i is the vehicle's own speed i∈{30,60,...,120}, and the yaw rate ω i Can be obtained from the vehicle CAN bus.

[0078]

[0079] Driving behavior data, also known as driving behavior descriptor, is a highly structured data model that aims to comprehensively reflect the driver's driving habits and vehicle dynamic performance in different scenarios, including the correspondence between the information of traffic participants around the vehicle and the vehicle's driving behavior parameters, that is, the correspondence between the type, distance and speed of traffic participants around the vehicle and the vehicle's longitudinal following distance statistics, acceleration and deceleration statistics, and lateral vehicle distance target distance statistics. In addition, the driving behavior data also includes the correspondence between the vehicle's speed range and the lane change yaw rate statistics, and / or the correspondence between the vehicle's speed range and the yaw rate change rate statistics.

[0080] Based on the calculation results of the driving behavior parameters of the above-mentioned vehicle, the driving behavior data of the vehicle is determined. In one embodiment, a driving behavior descriptor structure DBD can be constructed using methods such as neural networks, and its primary attributes include lane following descriptor CCD, lane keeping descriptor LKD and lane change descriptor LCD, and secondary attributes include following distance descriptor HWD, acceleration and deceleration descriptor ABD, steering descriptor SWD, lateral distance keeping descriptor LDD and reserved attribute descriptor. The specific acquisition method is as follows:

[0081] (1) Lane following descriptor CCD: It includes the following distance descriptor HWD and the acceleration / deceleration descriptor ABD. Both are stored in the form of a two-dimensional matrix (7x5), with the horizontal axis corresponding to the target category (1-sedan, 2-bus, 3-truck, 4-two-wheeler, 5-three-wheeler, 6-special vehicle, 7-other vehicles), and the vertical axis corresponding to the speed range (0-30, 30-60, 60-90, 90-120, 120+). HWD records the corresponding speed range and the average following distance hwd under the target vehicle category in the above two-dimensional table. i,j , ABD records the corresponding speed range and the average acceleration and deceleration amplitude a of the target vehicle category i,j .

[0082] (2) Lane keeping descriptor LKD: It is stored in the form of a two-dimensional matrix (8x5), with the horizontal axis corresponding to the target category (1-sedan, 2-bus, 3-truck, 4-two-wheeled vehicle, 5-three-wheeled vehicle, 6-special vehicle, 7-other vehicles, 8-road boundary), and the vertical axis corresponding to the speed range (0-30, 30-60, 60-90, 90-120, 120+). The above two-dimensional table records the corresponding speed range and the lateral target distance y under the target vehicle category in the adjacent lane. i,j (The target condition exists in one lane).

[0083] (3) Lane change descriptor LCD: The yaw rate ω of the lane change operation within the vehicle's own speed range (0-30, 30-60, 60-90, 90-120, 120+) i And the yaw rate change statistics describe.

[0084] (4) Reserved attribute descriptors: Some fields are reserved to backup and add other descriptors.

[0085] The driving behavior data of the vehicle determined in the above way integrates multi-dimensional driving behavior parameters in a highly structured way, providing a detailed and flexible data framework for driving behavior analysis, driver personalization evaluation, and learning and optimization of the autonomous driving system. This design not only facilitates the organization and analysis of data, but also promotes comparative research on driving behaviors across scenarios and categories.

[0086] S120: Training a driving model based on the driving behavior data of the vehicle.

[0087] The vehicle's driving model is used to describe and predict the vehicle's dynamic behavior under different driving operations and environmental conditions, reflecting the driving behavior preferences or personalized driving styles of different drivers. The driving model is trained based on the vehicle's driving behavior data to modify and update the vehicle control model parameters to match the driver's driving behavior preferences, making the assisted driving system more driver-friendly, thereby reducing the frequency of takeover and optimizing the user experience.

[0088] In one embodiment, the driving model training method is applied to a vehicle, and the vehicle sends driving behavior data to a cloud server so that the cloud server can use the driving behavior data to train the driving model.

[0089] In a specific embodiment, the driving behavior data of the vehicle is compressed and then communicated between the vehicle and the cloud using wireless communication methods such as 4G / 5G or other communication methods that can connect the vehicle-mounted terminal and the cloud server. In one example, a TCP / IP long link is used to ensure a stable two-way communication channel between the vehicle and the cloud; the vehicle can register with the cloud server using a user name and password and establish a verification mechanism; to maintain an active connection, the vehicle and the cloud server can maintain a connection between the two ends through a heartbeat method to confirm that the communication link is unobstructed.

[0090] At the same time, all data between the vehicle-mounted wireless module and the cloud server can be encrypted using a widely used symmetric encryption algorithm AES (Advanced Encryption Standard) such as AES128 or AES256 encryption algorithm to ensure data security during the communication process. Through this encryption method, even if the data is intercepted during transmission, it cannot be deciphered without the correct key, greatly reducing the risk of data leakage. In addition, combined with a secure authentication mechanism, the entire vehicle-cloud communication process can be carried out with high security, meeting the strict requirements of modern intelligent connected vehicles for data protection.

[0091] After the vehicle's driving behavior data is transmitted to the cloud server through vehicle-to-cloud communication, the cloud server generates a driving model through reinforcement learning training, including the auxiliary driving controller model parameters Gx and Gy. In one example, a highly accurate simulation platform such as CarSim or Simulink is used to simulate the vehicle's dynamic behavior and build a vehicle's dynamic model. The vehicle dynamic model takes into account a variety of physical parameters (Pv), such as mass, inertia, stiffness, etc. These parameters are obtained through vehicle calibration to determine the accuracy of the simulation. Dynamic state S t The update equation of is shown in the following formula (10), which reflects the vehicle response controller output U t After the state changes, the controller output

[0092] St+1 =VM(S t , P v , U t ) (10)

[0093] The controller model includes a longitudinal controller model XC and a lateral controller model YC, as shown in the following formulas (11) and (12), which are responsible for the acceleration / deceleration and steering control of the vehicle respectively. The controller output affects the longitudinal motion (such as speed control) and lateral motion (such as lane keeping) of the vehicle respectively. Among them, the key parameter of the controller G x , G y is the controller gain; S t is the vehicle dynamics state input to the controller at time t; TF t Traffic environment information, such as longitudinal and lateral target distance D x , D y , longitudinal and lateral adjacent target categories C x , C y , and lateral and longitudinal driving behavior category B x , B y wait.

[0094] [u t , a t ]=XC(G x ,S t ,TF t ) (11)

[0095]

[0096] Through reinforcement learning method, find the optimal controller gain G x , G y , allowing the vehicle to comply with safety regulations while imitating a specific driving style or achieving high efficiency and energy saving in various driving environments.

[0097] In one example, a reward / penalty function may also be defined. For example, a loss function is defined as shown in the following formula (13). By minimizing this loss function, the reinforcement learning algorithm continuously adjusts the controller parameters until the optimal G is found. x , G y combination.

[0098]

[0099] Among them, c(S t ,TF t ) = k 1 ·f(||TF t -E(TF)||)+k 2 ·g(||St -E(S)||)+k 3 collision, which represents the expected difference between the traffic environment description (following distance, lane keeping distance, etc.) and vehicle dynamics (acceleration, yaw rate, etc.) formed by the vehicle driving operation based on the control gain and the recorded driving behavior descriptor.

[0100] In one embodiment, the driving model training method is applied to a cloud server, the cloud server obtains the driving behavior data of the vehicle from the vehicle, and the cloud server trains the driving model based on the driving behavior data of the vehicle. The training method of the driving model of the cloud server is as described above. In addition, the cloud server can also use supervised learning, imitation learning and other methods to train the driving model, which will not be repeated here.

[0101] After training or optimizing the driving model based on the driving behavior data uploaded by the vehicle, the cloud server interacts with the vehicle to update the driving model on the vehicle to the trained driving model.

[0102] In one embodiment, the control model parameters of the vehicle are updated through the computing resources of the cloud server and reinforcement learning to complete the training and optimization of the driving model so that it matches the driver's driving preferences. Then, the model data is transmitted from the cloud server to the vehicle through a secure communication protocol to ensure the integrity and confidentiality of the model data during the transmission process to prevent the data from being intercepted or tampered with. In one example, the verified optimization model is packaged into a software package suitable for OTA upgrade, including version information, update logs, and security verification codes. An OTA upgrade notification is sent to the target vehicle through the cloud. After confirming receipt, the vehicle automatically downloads the update package and installs the new model at an appropriate time (such as when parking). After the vehicle installs the new model, self-diagnosis and functional verification are performed to ensure that the new model is installed correctly and can work properly. If there is an abnormality, the system will roll back to the previous version and report the error.

[0103] See also Figure 6 , Figure 6 is a flow chart of another embodiment of the driving model training method provided by the present application. It should be noted that if there are substantially the same results, this embodiment is not used. Figure 6 The process sequence shown is limited. Figure 6 As shown, this embodiment includes:

[0104] This application works in conjunction with cloud servers by making reasonable use of the edge computing power of the intelligent driving platform. Based on the correspondence between the traffic participant information perceived by the vehicle during driving and the driving behavior parameters, driving behavior data is generated. The driving behavior data is the result of the driving behavior description associated with the decision results for the strongly associated elements in the vehicle decision-making planning (such as traffic participant category, boundary type, and the vehicle's own driving status, etc.). The driving behavior data can more accurately record driving behavior preferences.

[0105] Based on vehicle-to-cloud communication, driving behavior data is uploaded to the cloud server, and the vehicle control model parameters are updated through cloud computing resources and reinforcement learning so that the driving model can match the driver's driving preferences. After the update, the data is pushed to the vehicle terminal through the over-the-air (OTA) upgrade technology to complete the replacement of the vehicle-side driving model.

[0106] In this embodiment, the process of edge computing and cloud collaboration in an intelligent connected vehicle can be specifically divided into two main stages: the edge fusion perception stage and the cloud driving style migration stage.

[0107] In the edge fusion perception stage, the on-board multimodal sensors play the role of information collection, covering different perception dimensions, including: (1) Visual perception: capturing image information of the vehicle's surrounding environment, identifying road signs, pedestrians and other vehicles, etc.; (2) Millimeter wave radar perception: sensing accurate distance and speed information, especially in severe weather conditions, which can supplement the deficiencies of visual perception; (3) Vehicle bus signal acquisition: reading the information of the vehicle's internal electronic control unit ECU, such as vehicle speed, accelerator pedal position and brake status, etc., to provide the vehicle's own status data.

[0108] Among them, if at least two perception methods are used to collect information about traffic participants, the visual perception and radar perception data can be fused to improve the accuracy and robustness of the perception system. The fused data can more accurately identify objects, estimate distances, and judge the movement trends of dynamic obstacles.

[0109] By combining the perception fusion results with the vehicle bus signals, the vehicle's driving behavior parameters are calculated, such as longitudinal following distance, acceleration and deceleration statistics, lateral vehicle distance target distance, lane change yaw rate, and yaw rate change rate. These parameters are the basis for evaluating and understanding driving style.

[0110] Based on the calculated driving behavior parameters, a driving behavior descriptor is constructed to obtain driving behavior data, that is, the correspondence between the information of traffic participants around the vehicle and the driving behavior parameters of the vehicle, specifically including the correspondence between the type, distance and speed of traffic participants around the vehicle and the longitudinal following distance statistics, acceleration and deceleration statistics, and lateral distance target distance statistics of the vehicle. In addition, the driving behavior data also includes the correspondence between the vehicle speed range and the lane change yaw rate statistics, and / or the correspondence between the vehicle speed range and the yaw rate change rate statistics, etc.

[0111] In order to efficiently transmit vehicle driving behavior data, the edge vehicle side compresses the data and transmits it to the cloud server through an encrypted channel, reducing communication costs and ensuring security and privacy protection during data transmission.

[0112] In the cloud-based driving style transfer phase, the cloud server receives the vehicle driving behavior data, decompresses and preprocesses it, and prepares it for driving model training. Based on the received driving behavior data, the cloud server uses a reinforcement learning algorithm to optimize the lateral and longitudinal control strategies of the autonomous driving controller. Among them, the adjustment of lateral and longitudinal gains such as acceleration and steering sensitivity is intended to simulate the style preferences of specific drivers and enhance the personalization and naturalness of the autonomous driving experience.

[0113] Finally, the optimized controller parameters are remotely updated to the vehicle-side autonomous driving controller through OTA (Over-The-Air) technology, realizing the migration of driving style and instant upgrade of control strategy.

[0114] See also Figure 7 , Figure 7 It is a structural diagram of an embodiment of a device provided in the present application. The device 60 includes a memory 61 and a processor 62 connected to each other. The memory 61 is used to store a computer program. When the computer program is executed by the processor 62, it is used to implement the driving model training method in the above embodiment.

[0115] The above-mentioned method is an embodiment, which can exist in the form of a computer program. Therefore, the present application proposes a computer-readable storage medium. Figure 8 , Figure 8 It is a structural diagram of an embodiment of a computer-readable storage medium provided in the present application. The computer-readable storage medium 80 is used to store a computer program 81, which can be executed to implement the driving model training method in the above embodiment.

[0116] The computer-readable storage medium 80 may be a server, a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or other media that can store program codes.

[0117] The above description is only an embodiment of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A driving model training method, It is characterized in that The method comprises: Determining driving behavior data of the vehicle, the driving behavior data including a correspondence between traffic participant information around the vehicle and a driving behavior parameter of the vehicle, the traffic participant information including a type and a speed of the traffic participant; A driving model is trained based on the driving behavior data of the vehicle.

2. The driving model training method according to claim 1, It is characterized in that The driving model training method is applied to the vehicle; The training of the driving model based on the driving behavior data of the vehicle includes: the vehicle sending the driving behavior data to a cloud server, so that the cloud server trains the driving model using the driving behavior data.

3. The driving model training method according to claim 2, It is characterized in that The determining of the driving behavior data of the vehicle comprises: Using at least two perception methods to detect the speed of the traffic participant, and obtaining at least two perceived speeds of the traffic participant; and / or using at least two perception methods to detect the distance of the traffic participant, and obtaining at least two perceived distances of the traffic participant; Based on the perception time difference between at least two perception methods, at least two perceived speeds of the traffic participant are fused to obtain the speed of the traffic participant; and / or, based on the perception time difference between at least two perception methods, at least two perceived distances of the traffic participant are fused to obtain the distance of the traffic participant.

4. The driving model training method according to claim 3, It is characterized in that The step of fusing at least two perceived speeds of the traffic participant based on the perceived time difference between at least two perception methods to obtain the speed of the traffic participant includes: Based on the Hungarian matching algorithm and the kinematic Kalman filter fusion, at least two perceived speeds of the traffic participant are fused to obtain the speed of the traffic participant; The step of fusing at least two perceived distances of the traffic participant based on the perceived time difference between at least two perception methods to obtain the distance of the traffic participant includes: Based on the Hungarian matching algorithm and the kinematic Kalman filter fusion, at least two perceived distances of the traffic participant are fused to obtain the distance of the traffic participant.

5. The driving model training method according to claim 3, It is characterized in that The detecting of the distance of the traffic participant by using at least two sensing methods includes: Analyzing video data around the vehicle using a video analysis network to obtain the type, speed and distance of the traffic participant in the video data; and / or, The radar signal collected by the vehicle is processed to obtain the distance and speed of the traffic participant.

6. The driving model training method according to claim 1, It is characterized in that The driving behavior parameters of the vehicle include longitudinal following distance statistics, acceleration and deceleration statistics, and lateral vehicle distance keeping target distance statistics; The driving behavior data further includes a correspondence between a speed range of the vehicle and a lane-changing yaw rate statistic, and / or a correspondence between a speed range of the vehicle and a yaw rate change rate statistic.

7. The driving model training method according to claim 1, It is characterized in that The driving model training method is applied to a cloud server, and the determining of the driving behavior data of the vehicle includes: the cloud server acquiring the driving behavior data of the vehicle from the vehicle; The training of the driving model based on the driving behavior data of the vehicle includes: the cloud server training the driving model based on the driving behavior data of the vehicle.

8. The driving model training method according to claim 7, It is characterized in that The cloud server trains a driving model based on the driving behavior data of the vehicle, and then includes: Interact with the vehicle to update the driving model on the vehicle to the trained driving model.

9. A device, It is characterized in that The device includes a processor and a memory, wherein the processor is coupled to the memory, and the processor is configured to execute one or more steps of the driving model training method according to any one of claims 1 to 8 based on instructions stored in the memory.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the driving model training method according to any one of claims 1 to 8.