A precise control method and system for key vehicle characteristics for ADAS test scenarios
By building and optimizing the vehicle key characteristic control model, the problem of insufficient vehicle control accuracy in ADAS testing is solved, high-precision control in complex scenarios is achieved, and the accuracy and reliability of test results are improved.
Patent Information
- Application Number
- CN202411300611.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-09-18
AI Technical Summary
Traditional vehicle control methods are difficult to meet the high requirements of Advanced Driver Assistance System (ADAS) testing on the control accuracy and consistency of key vehicle characteristics, especially in complex test scenarios, which affects the accuracy and reliability of test results.
By building state space and action space, collecting and learning trajectory operation data in different test scenarios, building trajectory operation reward functions, and building and optimizing the precise control model of key characteristics of the vehicle to achieve high key parameters such as vehicle speed and direction Accuracy, real-time control.
It realizes precise control of key vehicle characteristics in complex ADAS testing scenarios, improves the accuracy and reliability of test results, and ensures the safety and comfort of the vehicle in different driving scenarios.
Smart Images

Figure CN118819117B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of precise control of key vehicle characteristics, and in particular to a precise control method and system for key vehicle characteristics for ADAS test scenarios. Background Art
[0002] As an important development direction of the modern automobile industry, Advanced Driver Assistance Systems (ADAS) are rapidly becoming popular and developing. ADAS technology aims to improve driving safety, comfort and efficiency, including adaptive cruise control (ACC), automatic emergency braking (AEB), lane keeping assist (LKA) and other functions. With the continuous advancement and complexity of ADAS technology, its test and verification work has become more and more challenging. ADAS test scenarios usually involve the coordinated movement of multiple vehicles, and key characteristics such as the speed, position, and acceleration of the test vehicle need to be accurately controlled to simulate various complex traffic conditions. For example, in the AEB test, the motion trajectory and speed change of the target vehicle need to be accurately controlled to evaluate the braking performance of the vehicle under test. In these tests, the control accuracy of the key characteristics of the vehicle directly affects the accuracy and reliability of the test results. Traditional vehicle control methods mainly rely on manual driving or simple automatic control systems, which are difficult to meet the high requirements of ADAS testing for accuracy and consistency. In addition, with the increase and complexity of ADAS functions, the test scenarios have become more diverse and complex, which further increases the difficulty of vehicle control. Summary of the invention
[0003] In view of this, the present invention provides a method for accurately controlling key vehicle characteristics for ADAS test scenarios to adapt to various complex test scenarios and achieve high-precision, real-time control of key parameters such as vehicle speed and direction to support comprehensive testing and verification of ADAS functions.
[0004] To achieve the above object, the present invention provides a method for accurately controlling key characteristics of a vehicle for an ADAS test scenario, comprising the following steps:
[0005] S1: Construct the state space and action space of the test scenario, where the state space is represented by vectors based on the environmental state around the vehicle, and the action space is represented by vectors based on the key characteristic control parameters of the vehicle;
[0006] S2: Based on the constructed state space and action space, the trajectory operation data under different test scenarios are collected and represented by vectors to form a test trajectory operation data sequence. Preference learning is performed based on the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios.
[0007] S3: constructing a test scenario vehicle key characteristic precise control model based on the learned trajectory operation reward function, state space and action space, wherein the test scenario vehicle key characteristic precise control model takes the test scenario state as input and takes the vehicle key characteristic control parameter as output;
[0008] S4: Optimize and solve the constructed precise control model of key vehicle characteristics in the test scenario, and use the optimized model to precisely control the key vehicle characteristics in the test scenario.
[0009] As a further improvement method of the present invention:
[0010] Optionally, constructing the state space and action space of the test scenario in step S1 includes:
[0011] Construct the state space and action space of the test scenario, where the state space is represented by vectors based on the environmental state around the vehicle, and the action space is represented by vectors based on the key characteristic control parameters of the vehicle;
[0012] The state space uses sensors in the ADAS (Advanced Driving Assistance System) to collect the environmental state around the test vehicle, where the environmental state around the test vehicle is represented as follows:
[0013] ;
[0014] in:
[0015] C represents the representation form of the state of the environment surrounding the test vehicle;
[0016] Indicates the test vehicle's own position;
[0017] Indicates the environmental status of the eight directional areas around the test vehicle, where the environmental status of each directional area includes the position of the directional area, whether there is a vehicle, whether there is an obstacle, the vehicle's lateral speed, the vehicle's longitudinal speed, and the vehicle's direction;
[0018] The state of the environment around the test vehicle is encoded using a one-hot encoding method to obtain a state vector of the environment around the test vehicle;
[0019] The action space obtains the lateral speed, longitudinal speed and acceleration of the test vehicle as the key characteristic control parameters of the test vehicle, and encodes the key characteristic control parameters using a one-hot encoding method to obtain a key characteristic control parameter vector. In an embodiment of the present invention, the action space includes all the key characteristic control parameters of the test vehicle.
[0020] Optionally, in step S2, based on the constructed state space and action space, trajectory operation data under different test scenarios are collected and vectorized to form a test trajectory operation data sequence, including:
[0021] Based on the constructed state space and action space, the trajectory operation data of the test vehicle in different test scenarios are collected and represented by vectors to form a test trajectory operation data sequence, where the collected trajectory operation data sequence L is represented as:
[0022] ;
[0023] ;
[0024] in:
[0025] The vector representation result of the trajectory operation data in the nth test scenario;
[0026] The vector representation result representing the operation data of M consecutive road sections in the nth test scenario, The vector representing the operation data of the mth continuous road segment represents the result, The vector representation result representing the position of the test vehicle itself in the mth continuous road segment, The vector representation result representing the environmental state of the eight direction areas around the test vehicle in the mth continuous road section, A vector representation result representing the key characteristic control parameters of the test vehicle in the mth continuous road section;
[0027] Preference learning is performed based on the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios.
[0028] Optionally, performing preference learning according to the test trajectory operation data sequence to obtain trajectory operation reward functions under different test scenarios includes:
[0029] Preference learning is performed based on the test trajectory operation data sequence, and the trajectory operation reward function under different test scenarios is constructed. The construction process of the trajectory operation reward function under different test scenarios is as follows:
[0030] S21: Initialize the generated trajectory operation reward basis function:
[0031] ;
[0032] ;
[0033] ;
[0034] in:
[0035] represents the trajectory operation reward basis function, represents the state gap between the test vehicle and the preceding vehicle, the safety of the test vehicle, and the comfort of the test vehicle, respectively, and s represents the adjustment result of the key characteristic control parameters of the test vehicle;
[0036] Indicates the shortest distance between the test vehicle and surrounding vehicles under the current environment. It indicates the shortest distance between the test vehicle and surrounding vehicles after the test vehicle is adjusted according to the key characteristic control parameter adjustment result s;
[0037] Indicates the longitudinal speed of the vehicle ahead of the test vehicle under the current environmental conditions. It represents the longitudinal speed of the test vehicle after the test vehicle is adjusted according to the key characteristic control parameter adjustment result s;
[0038] Indicates the minimum vehicle distance threshold;
[0039] It indicates the change value of the vehicle acceleration of the test vehicle after the test vehicle is adjusted according to the key characteristic control parameter adjustment result s;
[0040] represents an exponential function with a natural constant as base;
[0041] S22: Obtaining trajectory operation data under different test scenarios, and calculating the preference between different test scenarios;
[0042] S23: taking the preference as the distance between different test scenarios, clustering the N test scenarios using the Kmeans algorithm to obtain K test scenario clusters of the test scenarios, wherein each test scenario cluster includes a number of test scenarios and a trajectory operation data vector corresponding to the test scenario, and the trajectory operation data vector is a vector representation result of the trajectory operation data;
[0043] S24: Calculate the trajectory operation reward function of each test scene cluster, where the trajectory operation reward function of the kth test scene cluster is:
[0044] ;
[0045] ;
[0046] ;
[0047] ;
[0048] in:
[0049] represents the trajectory operation reward function of the kth test scene cluster, ;
[0050] represents the mean vehicle distance, mean vehicle speed and mean acceleration of all trajectory operation data vectors in the kth test scenario cluster, where the mean vehicle distance represents the mean of the closest distance between the test vehicle and the surrounding vehicles in M sections, the mean vehicle speed represents the mean of the difference between the longitudinal speed of the test vehicle and the preceding vehicle in M sections, and the mean acceleration represents the mean acceleration of the test vehicle in M consecutive sections;
[0051] represents the vehicle distance standard deviation, vehicle speed standard deviation and acceleration standard deviation of all trajectory operation data vectors in the kth test scenario cluster;
[0052] Represents the weight values of the three trajectory operation reward basis functions in the kth test scene cluster.
[0053] Optionally, in step S22, the trajectory operation data under different test scenarios are obtained, and the preferences between the different test scenarios are calculated, including:
[0054] The nth test scenario is the same as the The preference between the test scenarios is:
[0055] ;
[0056] ;
[0057] in:
[0058] Indicates the relationship between the nth test scenario and the Preference between test scenarios, , ;
[0059] Representation vector With vector The similarity between represents the L2 norm.
[0060] Optionally, in step S3, a precise control model of key characteristics of the vehicle in the test scenario is constructed based on the learned trajectory operation reward function, state space, and action space, including:
[0061] Based on the learned trajectory operation reward function, state space and action space, a precise control model of key characteristics of the vehicle in the test scenario is constructed, wherein the precise control model of key characteristics of the vehicle in the test scenario takes the test scenario state as input and takes the key characteristic control parameters of the vehicle as output, wherein the improved deep Q network is an implementation method for constructing the precise control model of key characteristics of the vehicle in the test scenario, and the precise control model of key characteristics of the vehicle in the test scenario includes an input layer, an online network parameter extraction layer, a state mapping layer and an output layer;
[0062] The input layer is used to receive the test scene state, wherein the test scene state is the current test vehicle surrounding environment state of the test vehicle;
[0063] The online network parameter extraction layer is used to extract the network parameters deployed online and construct a neural network based on the network parameters;
[0064] The state mapping layer is used to map the test scene state using the constructed neural network to obtain the current vehicle key characteristic control parameters of the test vehicle as the output result of the output layer.
[0065] Optionally, the step S4 optimizes and solves the constructed test scenario vehicle key characteristic precise control model, including:
[0066] The key characteristics precise control model of the vehicle in the constructed test scenario is optimized and solved, and the optimization and solution process is as follows:
[0067] S41: Get the action space and initialize the generation model parameters ;
[0068] S42: Setting model parameters The current number of iterations is t, the maximum number of iterations is Max, and the initial value of t is 0, then the model parameter The result of the tth iteration is ;
[0069] S43: Select the trajectory operation data vectors of the first t test scenarios, and extract the environment state vector and the key characteristic control parameter vector therefrom, wherein the environment state vector is a vector representation of the surrounding environment state of the test vehicle at different road sections, and the key characteristic control parameter vector is a vector representation of the key characteristic control parameters of the test vehicle at different road sections;
[0070] S44: For the environmental state vector of any road section, use the model parameters Calculate the Q value of different key characteristic control parameter vectors in the action space for the environment state vector, select the key characteristic control parameter vector with the largest Q value as the output result, and calculate the trajectory operation reward function value of the key characteristic control parameter vector;
[0071] S45: Construct the mean square error loss function to obtain the trajectory operation reward function value:
[0072] ;
[0073] in:
[0074] represents the mean square error loss function;
[0075] T stands for transpose;
[0076] Indicates the selected The environmental feature vector corresponding to the mth road section of the test scene;
[0077] Represents the environmental feature vector The key characteristic control parameter vector with the largest corresponding Q value;
[0078] represents the L1 norm;
[0079] That is For the environmental feature vector Q value, b represents the control coefficient, Set to 0.2;
[0080] S46: Combined with the mean square error loss function, the model parameters are iteratively updated using the gradient descent method to obtain ;
[0081] S47: Let t = t + 1, return to step S44, until the maximum number of iterations is reached, and the optimal parameters are obtained , the network parameters Deploy online.
[0082] Optionally, in step S46, the mean square error loss function is combined to iteratively update the model parameters by using a gradient descent method, including:
[0083] S461: Calculate the mean square error loss function at the tth iteration Gradient :
[0084] ;
[0085] in:
[0086] represents a gradient operator; in the embodiment of the present invention, The calculation method is to use the model parameters as variables for the mean square error loss function Take the derivative and set the model parameters Substitute the derivative result;
[0087] S462: Calculate the moving average index :
[0088] ;
[0089] in:
[0090] represents the exponential decay rate, is set to 0.91; in the embodiment of the present invention, ;
[0091] S463: Iterate the model parameters:
[0092] ;
[0093] in:
[0094] Represents the learning rate, and the learning rate Set to 0.01.
[0095] Optionally, the method of using the model obtained by optimization to accurately control key characteristics of the vehicle in the test scenario includes:
[0096] The model obtained by optimization is used to accurately control the key characteristics of the vehicle in the test scenario. The process of accurately controlling the key characteristics of the vehicle in the test scenario is as follows:
[0097] The input layer receives the test scenario state S of the test vehicle;
[0098] The online network parameter extraction layer extracts the network parameters deployed online , and based on network parameters A neural network is constructed, which includes a convolutional layer and a fully connected layer. The network parameters The weight matrix of the convolutional layer is contained in And the mapping matrix of the fully connected layer , ;
[0099] The state mapping layer uses the constructed neural network to map the test scene state S and obtains the current vehicle key characteristic control parameter A of the test vehicle as the output result of the output layer:
[0100] ;
[0101] in:
[0102] T stands for transpose;
[0103] It represents the inverse encoding of the one-hot encoding, converts the key characteristic control parameter vector into the key characteristic control parameter, realizes the use of ADAS to collect environmental status, and generates key characteristic control parameters to ensure vehicle safety and comfort and reduce the correlation with the distance and speed of the leading vehicle.
[0104] In order to solve the above problems, the present invention provides a vehicle key characteristic precision control system for ADAS test scenarios, characterized in that the system comprises:
[0105] A test data acquisition module is used to construct the state space and action space of the test scenario, collect the trajectory operation data under different test scenarios based on the constructed state space and action space, and perform vector representation to form a test trajectory operation data sequence;
[0106] The function generation module is used to perform preference learning based on the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios;
[0107] The vehicle control device is used to construct a precise control model of the key characteristics of the vehicle in the test scenario based on the learned trajectory operation reward function, state space and action space, optimize and solve the constructed precise control model of the key characteristics of the vehicle in the test scenario, and use the model obtained by the optimization solution to accurately control the key characteristics of the vehicle in the test scenario.
[0108] In order to solve the above problem, the present invention further provides an electronic device, the electronic device comprising:
[0109] A memory storing at least one instruction;
[0110] Communication interface, enabling electronic equipment to communicate; and
[0111] The processor executes the instructions stored in the memory to implement the above-mentioned vehicle key characteristic precise control method for ADAS test scenarios.
[0112] In order to solve the above problems, the present invention also provides a computer-readable storage medium, in which at least one instruction is stored. The at least one instruction is executed by a processor in an electronic device to implement the above-mentioned method for precise control of key vehicle characteristics for ADAS test scenarios.
[0113] Compared with the prior art, the present invention proposes a method for accurately controlling key vehicle characteristics for ADAS test scenarios, which has the following advantages:
[0114] First, this scheme proposes a reward function construction method that combines scene characteristics. It uses ADAS to collect environmental state information and key characteristic control parameter information in different scenes to form state space and action space, and performs preference calculation and clustering processing on different test scenes. According to the clustering results of different test scenes for driving safety, comfort and the degree of association with the preceding vehicle, it generates trajectory operation reward functions in different test scenes, realizes vehicle characteristic adjustment based on scene characteristics, and focuses on different driving experiences in different driving scenes.
[0115] At the same time, this scheme quantifies the optimization effect of model parameters in the precise control model of key characteristics of the vehicle in the test scenario based on the trajectory operation reward function, and synchronously iterates the model parameters by iterating the number of training data. The model obtained by optimization is used to receive the test scene status collected by ADAS, and the model parameters are used for mapping to generate key characteristic control parameters that ensure vehicle safety, comfort, and reduce the distance and speed correlation with the vehicle in front. BRIEF DESCRIPTION OF THE DRAWINGS
[0116] Figure 1 A schematic flow chart of a method for accurately controlling key vehicle characteristics in an ADAS test scenario provided by one embodiment of the present invention;
[0117] Figure 2 A functional module diagram of a vehicle key characteristic precision control system for ADAS test scenarios provided by an embodiment of the present invention;
[0118] Figure 2 In: 100 precise control system of key vehicle characteristics for ADAS test scenarios, 101 test data acquisition module, 102 function generation module, 103 vehicle control device;
[0119] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0120] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.
[0121] The embodiment of the present application provides a method for accurately controlling key characteristics of a vehicle for an ADAS test scenario. The execution subject of the method for accurately controlling key characteristics of a vehicle for an ADAS test scenario includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the method for accurately controlling key characteristics of a vehicle for an ADAS test scenario can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc.
[0122] S1: Construct the state space and action space of the test scenario, where the state space is represented by vectors based on the environmental state around the vehicle, and the action space is represented by vectors based on the key characteristic control parameters of the vehicle.
[0123] The state space and action space of the test scenario are constructed in step S1, including:
[0124] Construct the state space and action space of the test scenario, where the state space is represented by vectors based on the environmental state around the vehicle, and the action space is represented by vectors based on the key characteristic control parameters of the vehicle;
[0125] The state space uses sensors in the ADAS (Advanced Driving Assistance System) to collect the environmental state around the test vehicle, where the environmental state around the test vehicle is represented as follows:
[0126] ;
[0127] in:
[0128] C represents the representation form of the state of the environment surrounding the test vehicle;
[0129] Indicates the test vehicle's own position;
[0130] Indicates the environmental status of the eight directional areas around the test vehicle, where the environmental status of each directional area includes the position of the directional area, whether there is a vehicle, whether there is an obstacle, the vehicle's lateral speed, the vehicle's longitudinal speed, and the vehicle's direction;
[0131] The state of the environment around the test vehicle is encoded using a one-hot encoding method to obtain a state vector of the environment around the test vehicle;
[0132] The action space obtains the lateral velocity, longitudinal velocity and acceleration of the test vehicle itself as the key characteristic control parameters of the test vehicle, and encodes the key characteristic control parameters using the one-hot encoding method to obtain the key characteristic control parameter vector.
[0133] S2: Based on the constructed state space and action space, the trajectory operation data under different test scenarios are collected and represented by vectors to form a test trajectory operation data sequence. Preference learning is performed based on the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios.
[0134] In the step S2, based on the constructed state space and action space, the trajectory operation data under different test scenarios are collected and vectorized to form a test trajectory operation data sequence, including:
[0135] Based on the constructed state space and action space, the trajectory operation data of the test vehicle in different test scenarios are collected and represented by vectors to form a test trajectory operation data sequence, where the collected trajectory operation data sequence L is represented as:
[0136] ;
[0137] ;
[0138] in:
[0139] The vector representation result of the trajectory operation data in the nth test scenario;
[0140] The vector representation result representing the operation data of M consecutive road sections in the nth test scenario, The vector representing the operation data of the mth continuous road segment represents the result, The vector representation result representing the position of the test vehicle itself in the mth continuous road segment, The vector representation result representing the environmental state of the eight direction areas around the test vehicle in the mth continuous road section, A vector representation result representing the key characteristic control parameters of the test vehicle in the mth continuous road section;
[0141] Preference learning is performed based on the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios.
[0142] The preference learning is performed according to the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios, including:
[0143] Preference learning is performed based on the test trajectory operation data sequence, and the trajectory operation reward function under different test scenarios is constructed. The construction process of the trajectory operation reward function under different test scenarios is as follows:
[0144] S21: Initialize the generated trajectory operation reward basis function:
[0145] ;
[0146] ;
[0147] ;
[0148] in:
[0149] represents the trajectory operation reward basis function, represents the state gap between the test vehicle and the preceding vehicle, the safety of the test vehicle, and the comfort of the test vehicle, respectively, and s represents the adjustment result of the key characteristic control parameters of the test vehicle;
[0150] Indicates the shortest distance between the test vehicle and surrounding vehicles under the current environment. It indicates the shortest distance between the test vehicle and surrounding vehicles after the test vehicle is adjusted according to the key characteristic control parameter adjustment result s;
[0151] Indicates the longitudinal speed of the vehicle ahead of the test vehicle under the current environmental conditions. It represents the longitudinal speed of the test vehicle after the test vehicle is adjusted according to the key characteristic control parameter adjustment result s;
[0152] Indicates the minimum vehicle distance threshold;
[0153] It indicates the change value of the vehicle acceleration of the test vehicle after the test vehicle is adjusted according to the key characteristic control parameter adjustment result s;
[0154] represents an exponential function with a natural constant as base;
[0155] S22: Obtaining trajectory operation data under different test scenarios, and calculating the preference between different test scenarios;
[0156] S23: taking the preference as the distance between different test scenarios, clustering the N test scenarios using the Kmeans algorithm to obtain K test scenario clusters of the test scenarios, wherein each test scenario cluster includes a number of test scenarios and a trajectory operation data vector corresponding to the test scenario, and the trajectory operation data vector is a vector representation result of the trajectory operation data;
[0157] S24: Calculate the trajectory operation reward function of each test scene cluster, where the trajectory operation reward function of the kth test scene cluster is:
[0158] ;
[0159] ;
[0160] ;
[0161] ;
[0162] in:
[0163] represents the trajectory operation reward function of the kth test scene cluster, ;
[0164] represents the mean vehicle distance, mean vehicle speed and mean acceleration of all trajectory operation data vectors in the kth test scenario cluster, where the mean vehicle distance represents the mean of the closest distance between the test vehicle and the surrounding vehicles in M sections, the mean vehicle speed represents the mean of the difference between the longitudinal speed of the test vehicle and the preceding vehicle in M sections, and the mean acceleration represents the mean acceleration of the test vehicle in M consecutive sections;
[0165] represents the vehicle distance standard deviation, vehicle speed standard deviation and acceleration standard deviation of all trajectory operation data vectors in the kth test scenario cluster;
[0166] Represents the weight values of the three trajectory operation reward basis functions in the kth test scene cluster.
[0167] In the step S22, the trajectory operation data under different test scenarios are obtained, and the preference between different test scenarios is calculated, including:
[0168] The nth test scenario is the same as the The preference between the test scenarios is:
[0169] ;
[0170] ;
[0171] in:
[0172] Indicates the relationship between the nth test scenario and the Preference between test scenarios, , ;
[0173] Representation vector With vector The similarity between represents the L2 norm.
[0174] S3: Construct a precise control model of key characteristics of the vehicle in the test scenario based on the learned trajectory operation reward function, state space and action space. The precise control model of key characteristics of the vehicle in the test scenario takes the test scenario state as input and takes the key characteristics control parameters of the vehicle as output.
[0175] In the step S3, a precise control model of key characteristics of the vehicle in the test scenario is constructed based on the learned trajectory operation reward function, state space, and action space, including:
[0176] Based on the learned trajectory operation reward function, state space and action space, a test scenario vehicle key characteristic precise control model is constructed, wherein the test scenario vehicle key characteristic precise control model takes the test scenario state as input and takes the vehicle key characteristic control parameter as output, and the test scenario vehicle key characteristic precise control model includes an input layer, an online network parameter extraction layer, a state mapping layer and an output layer;
[0177] The input layer is used to receive the test scene state, wherein the test scene state is the current test vehicle surrounding environment state of the test vehicle;
[0178] The online network parameter extraction layer is used to extract the network parameters deployed online and construct a neural network based on the network parameters;
[0179] The state mapping layer is used to map the test scene state using the constructed neural network to obtain the current vehicle key characteristic control parameters of the test vehicle as the output result of the output layer.
[0180] S4: Optimize and solve the constructed precise control model of key vehicle characteristics in the test scenario, and use the optimized model to precisely control the key vehicle characteristics in the test scenario.
[0181] The S4 step optimizes and solves the constructed test scenario vehicle key characteristic precise control model, including:
[0182] The key characteristics precise control model of the vehicle in the constructed test scenario is optimized and solved, and the optimization and solution process is as follows:
[0183] S41: Get the action space and initialize the generation model parameters ;
[0184] S42: Setting model parameters The current number of iterations is t, the maximum number of iterations is Max, and the initial value of t is 0, then the model parameter The result of the tth iteration is ;
[0185] S43: Select the trajectory operation data vectors of the first t test scenarios, and extract the environment state vector and the key characteristic control parameter vector therefrom, wherein the environment state vector is a vector representation of the surrounding environment state of the test vehicle at different road sections, and the key characteristic control parameter vector is a vector representation of the key characteristic control parameters of the test vehicle at different road sections;
[0186] S44: For the environmental state vector of any road section, use the model parameters Calculate the Q value of different key characteristic control parameter vectors in the action space for the environment state vector, select the key characteristic control parameter vector with the largest Q value as the output result, and calculate the trajectory operation reward function value of the key characteristic control parameter vector;
[0187] S45: Construct the mean square error loss function to obtain the trajectory operation reward function value:
[0188] ;
[0189] in:
[0190] represents the mean square error loss function;
[0191] T stands for transpose;
[0192] Indicates the selected The environmental feature vector corresponding to the mth road section of the test scene;
[0193] Represents the environmental feature vector The key characteristic control parameter vector with the largest corresponding Q value;
[0194] represents the L1 norm;
[0195] S46: Combined with the mean square error loss function, the model parameters are iteratively updated using the gradient descent method to obtain ;
[0196] S47: Let t = t + 1, return to step S44, until the maximum number of iterations is reached, and the optimal parameters are obtained , the network parameters Deploy online.
[0197] In step S46, the mean square error loss function is combined to iteratively update the model parameters using a gradient descent method, including:
[0198] S461: Calculate the mean square error loss function at the tth iteration Gradient :
[0199] ;
[0200] in:
[0201] represents a gradient operator; in the embodiment of the present invention, The calculation method is to use the model parameters as variables for the mean square error loss function Take the derivative and set the model parameters Substitute the derivative result;
[0202] S462: Calculate the moving average index :
[0203] ;
[0204] in:
[0205] represents the exponential decay rate, is set to 0.91; in the embodiment of the present invention, ;
[0206] S463: Iterate the model parameters:
[0207] ;
[0208] in:
[0209] Represents the learning rate, and the learning rate Set to 0.01.
[0210] The model obtained by optimization is used to accurately control the key characteristics of the vehicle in the test scenario, including:
[0211] The model obtained by optimization is used to accurately control the key characteristics of the vehicle in the test scenario. The process of accurately controlling the key characteristics of the vehicle in the test scenario is as follows:
[0212] The input layer receives the test scenario state S of the test vehicle;
[0213] The online network parameter extraction layer extracts the network parameters deployed online , and based on network parameters A neural network is constructed, which includes a convolutional layer and a fully connected layer. The network parameters The weight matrix of the convolutional layer is contained in And the mapping matrix of the fully connected layer , ;
[0214] The state mapping layer uses the constructed neural network to map the test scene state S and obtains the current vehicle key characteristic control parameter A of the test vehicle as the output result of the output layer:
[0215] ;
[0216] in:
[0217] T stands for transpose;
[0218] The inverse encoding of the one-hot encoding is used to convert the key characteristic control parameter vector into the key characteristic control parameter, so as to collect the environmental status using ADAS and generate the key characteristic control parameters to ensure the safety and comfort of the vehicle and reduce the correlation with the distance and speed of the vehicle in front.
[0219] like Figure 2 As shown, it is a functional module diagram of a vehicle key characteristic precise control system for ADAS test scenarios provided by an embodiment of the present invention, which can implement the vehicle key characteristic precise control method for ADAS test scenarios in Example 1.
[0220] The vehicle key characteristic precision control system 100 for ADAS test scenarios of the present invention can be installed in an electronic device. According to the functions implemented, the vehicle key characteristic precision control system for ADAS test scenarios can include a test data acquisition module 101, a function generation module 102 and a vehicle control device 103. The module of the present invention can also be called a unit, which refers to a series of computer program segments that can be executed by an electronic device processor and can complete fixed functions, which are stored in the memory of the electronic device.
[0221] The test data collection module 101 is used to construct the state space and action space of the test scenario, collect the trajectory operation data under different test scenarios based on the constructed state space and action space, and perform vector representation to form a test trajectory operation data sequence;
[0222] A function generation module 102 is used to perform preference learning according to the test trajectory operation data sequence to obtain trajectory operation reward functions under different test scenarios;
[0223] The vehicle control device 103 is used to construct a precise control model of the key characteristics of the vehicle in the test scenario based on the learned trajectory operation reward function, state space and action space, optimize and solve the constructed precise control model of the key characteristics of the vehicle in the test scenario, and use the model obtained by the optimization solution to accurately control the key characteristics of the vehicle in the test scenario.
[0224] In detail, each module in the vehicle key characteristic precision control system 100 for ADAS test scenarios in the embodiment of the present invention is used in the same manner as above. Figure 1 The same technical means are used as the method for precise control of key vehicle characteristics for ADAS test scenarios described in the previous section and can produce the same technical effects, so they will not be repeated here.
[0225] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.
[0226] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.
[0227] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A method for accurately controlling key vehicle characteristics for ADAS test scenarios, characterized in that: The method comprises: S1: Construct the state space and action space of the test scenario, where the state space is represented by vectors based on the environmental state around the vehicle, and the action space is represented by vectors based on the key characteristic control parameters of the vehicle; S2: Based on the constructed state space and action space, the trajectory operation data under different test scenarios are collected and represented by vectors to form a test trajectory operation data sequence. Preference learning is performed according to the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios, including: obtaining the trajectory operation data under different test scenarios, and calculating the preference between different test scenarios, where: The nth test scenario is the same as the The preference between the test scenarios is: ; ; in: Indicates the relationship between the nth test scenario and the Preference between test scenarios, , ; Representation vector With vector The similarity between represents the L2 norm; The vector representation result representing the operation data of the mth continuous road segment under the nth test scenario; Indicates The vector representation result of the operation data of the pth continuous road section in the test scenario; S3: constructing a test scenario vehicle key characteristic precise control model based on the learned trajectory operation reward function, state space and action space, wherein the test scenario vehicle key characteristic precise control model takes the test scenario state as input and takes the vehicle key characteristic control parameter as output; S4: Optimize and solve the constructed precise control model of key vehicle characteristics in the test scenario, and use the optimized model to precisely control the key vehicle characteristics in the test scenario.
2. A method for accurately controlling key vehicle characteristics for ADAS test scenarios as claimed in claim 1, characterized in that: The state space and action space of the test scenario are constructed in step S1, including: Construct the state space and action space of the test scenario, where the state space is represented by vectors based on the environmental state around the vehicle, and the action space is represented by vectors based on the key characteristic control parameters of the vehicle; The state space collects the environmental state around the test vehicle, where the environmental state around the test vehicle is represented as: ; in: C represents the representation form of the state of the environment surrounding the test vehicle; Indicates the test vehicle's own position; Indicates the environmental status of the eight directional areas around the test vehicle, where the environmental status of each directional area includes the position of the directional area, whether there is a vehicle, whether there is an obstacle, the vehicle's lateral speed, the vehicle's longitudinal speed, and the vehicle's direction. Indicates the area directly ahead. Indicates the area directly behind. Indicates the left area, Indicates the right area, Indicates the left front area, Indicates the right front area, Represents the left rear area, Indicates the right rear area; The state of the environment around the test vehicle is encoded using a one-hot encoding method to obtain a state vector of the environment around the test vehicle; The action space obtains the lateral velocity, longitudinal velocity and acceleration of the test vehicle itself as the key characteristic control parameters of the test vehicle, and encodes the key characteristic control parameters using the one-hot encoding method to obtain the key characteristic control parameter vector.
3. A vehicle key characteristic precision control method for ADAS test scenarios as claimed in claim 2, characterized in that: In the step S2, based on the constructed state space and action space, the trajectory operation data under different test scenarios are collected and vectorized to form a test trajectory operation data sequence, including: Based on the constructed state space and action space, the trajectory operation data of the test vehicle in different test scenarios are collected and represented by vectors to form a test trajectory operation data sequence, where the collected trajectory operation data sequence L is represented as: ; ; in: The vector representation result of the trajectory operation data in the nth test scenario; The vector representation result representing the operation data of M consecutive road sections in the nth test scenario, The vector representing the operation data of the mth continuous road segment represents the result, The vector representation result representing the position of the test vehicle itself in the mth continuous road segment, The vector representation result representing the environmental state of the eight direction areas around the test vehicle in the mth continuous road section, A vector representation result representing the key characteristic control parameters of the test vehicle in the mth continuous road section; Preference learning is performed based on the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios.
4. A method for accurately controlling key vehicle characteristics for ADAS test scenarios as claimed in claim 3, characterized in that: The preference learning is performed according to the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios, including: S21: Initialize the generated trajectory operation reward basis function; S22: Obtaining trajectory operation data under different test scenarios, and calculating the preference between different test scenarios; S23: taking the preference as the distance between different test scenarios, clustering the N test scenarios using the Kmeans algorithm to obtain K test scenario clusters of the test scenarios, wherein each test scenario cluster includes a number of test scenarios and a trajectory operation data vector corresponding to the test scenario, and the trajectory operation data vector is a vector representation result of the trajectory operation data; S24: Calculate the trajectory operation reward function for each test scene cluster.
5. A method for accurately controlling key vehicle characteristics for ADAS test scenarios as claimed in claim 1, characterized in that: In the step S3, a precise control model of key characteristics of the vehicle in the test scenario is constructed based on the learned trajectory operation reward function, state space, and action space, including: Based on the learned trajectory operation reward function, state space and action space, a test scenario vehicle key characteristic precise control model is constructed, wherein the test scenario vehicle key characteristic precise control model takes the test scenario state as input and takes the vehicle key characteristic control parameter as output, and the test scenario vehicle key characteristic precise control model includes an input layer, an online network parameter extraction layer, a state mapping layer and an output layer; The input layer is used to receive the test scene state, wherein the test scene state is the current test vehicle surrounding environment state of the test vehicle; The online network parameter extraction layer is used to extract the network parameters deployed online and construct a neural network based on the network parameters; The state mapping layer is used to map the test scene state using the constructed neural network to obtain the current vehicle key characteristic control parameters of the test vehicle as the output result of the output layer.
6. A method for accurately controlling key vehicle characteristics for ADAS test scenarios as claimed in claim 5, characterized in that: The S4 step optimizes and solves the constructed test scenario vehicle key characteristic precise control model, including: The key characteristics precise control model of the vehicle in the constructed test scenario is optimized and solved, and the optimization and solution process is as follows: S41: Obtain the action space and initialize the generation model parameters; S42: Set the current iteration number of the model parameter to t, the maximum iteration number to Max, and the initial value of t to 0. Then the tth iteration result of the model parameter is ; S43: extracting an environment state vector and a key characteristic control parameter vector from the trajectory operation data of the test scene, wherein the environment state vector is a vector representation of the surrounding environment state of the test vehicle at different road sections, and the key characteristic control parameter vector is a vector representation of the key characteristic control parameters of the test vehicle at different road sections; S44: for the environment state vector of any road section, the Q value of different key characteristic control parameter vectors in the action space for the environment state vector is calculated using the model parameters, the key characteristic control parameter vector with the largest Q value is selected as the output result, and the trajectory operation reward function value of the key characteristic control parameter vector is calculated; S45: construct a mean square error loss function to obtain the trajectory operation reward function value; S46: Combined with the mean square error loss function, the model parameters are iteratively updated using the gradient descent method; S47: Let t = t + 1, return to step S44, until the maximum number of iterations is reached, and the optimal model parameters are obtained , the network parameters Deploy online.
7. A method for accurately controlling key vehicle characteristics for ADAS test scenarios as claimed in claim 6, characterized in that: In step S46, the mean square error loss function is combined to iteratively update the model parameters using a gradient descent method, including: S461: Calculate the gradient of the mean square error loss function at the t-th iteration; S462: Update and calculate the moving average index of the current round; S463: Iteratively optimize the model parameters according to the calculated moving average index of the current round.
8. A vehicle key characteristic precision control method for ADAS test scenarios as claimed in claim 1, characterized in that: The model obtained by optimization is used to accurately control the key characteristics of the vehicle in the test scenario, including: The model obtained by optimization is used to accurately control the key characteristics of the vehicle in the test scenario. The process of accurately controlling the key characteristics of the vehicle in the test scenario is as follows: The input layer receives the test scenario state S of the test vehicle; The online network parameter extraction layer extracts the network parameters deployed online , and based on network parameters A neural network is constructed, wherein the constructed neural network includes a convolutional layer and a fully connected layer; The state mapping layer uses the constructed neural network to map the test scene state S and obtains the current vehicle key characteristic control parameter A of the test vehicle as the output result of the output layer.
9. A precise control system for key vehicle characteristics for ADAS test scenarios, characterized in that: The system comprises: A test data acquisition module is used to construct the state space and action space of the test scenario, collect the trajectory operation data under different test scenarios based on the constructed state space and action space, and perform vector representation to form a test trajectory operation data sequence; The function generation module is used to perform preference learning based on the test trajectory operation data sequence to obtain the trajectory operation reward function under different test scenarios; A vehicle control device, for constructing a precise control model of key characteristics of a vehicle in a test scenario based on the learned trajectory operation reward function, state space, and action space, optimizing and solving the constructed precise control model of key characteristics of the vehicle in the test scenario, and using the model obtained by the optimization solution to precisely control the key characteristics of the vehicle in the test scenario, so as to realize the precise control method of key characteristics of a vehicle for an ADAS test scenario as described in any one of claims 1 to 8.
