Automatic driving system test coverage rate evaluation method based on scene representativeness

Abstract

The invention relates to an automatic driving system test coverage rate evaluation method based on scene representativeness. Comprising the following steps: step 1, defining a logic scene parameter space and sampling to obtain a specific scene parameter set S; 2, establishing a scene representativeness quantitative evaluation method; and step 3, carrying out quantitative evaluation on the test coverage rate. According to the invention, through setting a scene representativeness quantitative evaluation index, quantitative evaluation is carried out on the coverage of a small number of specific scenes obtained by sampling in a logic scene parameter space by an enterprise or a third-party detection mechanism in the logic scene parameter space. Besides, on one hand, the coverage rate of the test process can be evaluated, and on the other hand, an enterprise can be guided to sample and generate a full-coverage test scene point set in the parameter space, and a standardized test scene combination under the scene is established. An enterprise or a third-party detection mechanism can be effectively helped to establish a standardized self-driving automobile test process, and the performance verification process of the self-driving automobile is promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving vehicle testing, and specifically provides a method for evaluating the test coverage of an autonomous driving system based on scenario representativeness. Background Art

[0002] Scenario-based testing methods are the mainstream means of verifying the safety of autonomous driving systems. As the main body of scenario library construction, logical scenarios use parameter spaces to describe the parameters of scenario elements. In theory, an infinite number of scenarios can be sampled in a continuous space. However, how to describe the coverage of the logical scenario parameter space by a small number of samples sampled by enterprises or third-party testing agencies has not been effectively solved. Existing scholars have conducted some research on scenario coverage, such as the patent "An accelerated test method for autonomous driving that considers efficiency and coverage" (patent number CN202111398403.2) and the flexible test board for improving the sensor I / O coverage of autonomous driving platforms (CN201911258752.7). However, most existing methods only focus on the test scenario generation process, that is, guiding the generation of test scenarios from the perspective of improving the comprehensiveness of the test, and cannot solve the problem of quantitative evaluation of the coverage of the parameter space by the generated test scenarios. Summary of the Invention

[0003] To address the above-mentioned issues, the present invention provides a method for evaluating the test coverage of an autonomous driving system based on scenario representativeness. By setting a quantitative evaluation index for scenario representativeness, the coverage of a small number of specific scenarios sampled by an enterprise or a third-party testing agency in the logical scenario parameter space is quantitatively evaluated. In addition, the present invention can, on the one hand, evaluate the coverage of the test process, and on the other hand, guide enterprises to sample and generate a set of fully covered test scenario points in the parameter space, and establish a standardized test scenario combination under the scenario. This can effectively help enterprises or third-party testing agencies establish a standardized autonomous driving vehicle testing process and promote the autonomous driving vehicle performance verification process.

[0004] The technical solution of the present invention is described as follows in conjunction with the accompanying drawings:

[0005] The present invention provides a method for evaluating test coverage of an autonomous driving system based on scenario representativeness, comprising the following steps:

[0006] Step 1: Define the logical scene parameter space and sample the specific scene parameter set S;

[0007] Step 2: Establish a quantitative evaluation method for scene representativeness;

[0008] Step 3: Quantitatively evaluate the test coverage.

[0009] Furthermore, the specific method of step 2 is as follows:

[0010] 21) In the logical scenario parameter space, the representativeness of a specific scenario is determined by two inherent properties: probability and risk. Probability represents the frequency of a scenario in real driving, while risk represents the accident situation of a scenario in a real driving environment. Therefore, a scenario representativeness quantification function driven by both risk and probability is designed to adaptively determine the radiation range of each specific scenario in the parameter space, as follows:

[0011]

[0012] Where r i (j) is the radiation range of the i-th specific scene in the j-th dimension of the parameter space, that is, representativeness; i is the scene danger index of the i-th scene; A i is the scenario probability index of the i-th specific scenario; λ is the risk factor adjustment index; δ is the scaling constant of the Gaussian probability density function; ε is the probability adjustment constant; β is the global influence range control coefficient; ω j is the weight coefficient of the j-th dimension;

[0013] In formula (1), the logical scene parameter space is discretized and r is limited. i (j) The influence range is rounded down to an integer; at the same time, the minimum value of the global influence range is set to r min Prevent point-like impact areas; set the maximum value r max Prevent the range from expanding infinitely; set (ξ i / maxξ) λ , used to normalize the risk level parameters; introduce independent weight coefficients ω for each dimension into the adaptive function of the test point influence range to adjust the expansion capabilities in different directions;

[0014] 22) For the scenario probability index A i , modeled using a Gaussian distribution as follows:

[0015]

[0016] Where g is the parameter dimension; ∑ is the covariance matrix used to describe the correlation between various types of parameters; μ is the mean vector of each parameter variable; X is the parameter vector of the specific scenario;

[0017] 23) For the scene danger index ξ i, considering both collision and non-collision scenarios for calculation; for collision scenarios, it is divided into three categories according to the relative speed of the two vehicles at the time of collision: severe collision, general collision and minor collision. Severe collision means that the relative speed of the two vehicles at the time of collision is greater than 65km / h, general collision is 20km / h to 65km / h, and minor collision is less than 20km / h; for non-collision scenarios, A is calculated based on the inverse of the collision time, i.e., Time to Collision, TTC. TTC -1 The calculation is as follows:

[0018]

[0019] Where Δdis is the distance between the front and rear vehicles; Δv is the relative speed between the front and rear vehicles; 0.7s -1 TTC as a near-collision and safety scenario -1 boundary;

[0020] Taking into account both collision and non-collision situations, the scene hazard index ξ i It is defined as a five-level risk rating, with severe collision at 1.0, general collision at 0.8, minor collision at 0.6, near collision at 0.4, and safety scenario at 0.2;

[0021] Use the proxy model to infer the unknown scene danger index, and then obtain the danger index of different specific scenes;

[0022] The driver agent model is divided into two parts: longitudinal and lateral. The longitudinal agent model is modeled using a long short-term memory network, while the lateral agent model is modeled using a lane-changing potential energy model and Monte Carlo tree search.

[0023] For the lateral agent model, the lane-changing potential model quantifies the driver's satisfaction with the current driving environment and provides dynamic intention input for lane-changing planning; MCTS explores the optimal lane-changing strategy by simulating multiple possible lane-changing paths.

[0024] The lane-changing potential energy model is defined by comprehensively considering vehicle speed, mass, relative information of the preceding vehicle, and road adhesion conditions. The calculation formula is shown in (4), and the U threshold for generating lane-changing intention is defined as 0.15;

[0025]

[0026] Where I is a constant used to adjust the magnitude of potential energy and is set to 0.001; M i is the equivalent mass of the target vehicle, which is 5000 kg; r ij is the relative distance between the main vehicle and the target vehicle; Δv is the relative speed; θ iThe angle between the speed direction of the main vehicle and the direction of the line connecting the target vehicle; l1 and l2 are attenuation factors, which are used to describe the exponential effects of distance and speed on potential energy, and are set to 1 and 0.05 respectively;

[0027] In formula (4), the environmental perception weight W is adjusted according to the road adhesion coefficient under different driving environments. Under low adhesion coefficient conditions, the driver will actively reduce the vehicle dynamic response, thereby slowing down the accumulation of potential energy; as shown in formula (5):

[0028]

[0029] Where ρ is a constant used to adjust the influence of environmental perception weight; o is the road adhesion coefficient;

[0030] After generating the lane-changing intention, MCTS is used to search for the optimal lane-changing path.

[0031] Furthermore, the specific method of the MCTS to search for the optimal lane change path is as follows:

[0032] (a) selection;

[0033] Starting from the root node, select an optimal child node along the Monte Carlo tree structure; the goal of the selection is to maximize the value of the decision strategy;

[0034] (b) expansion;

[0035] During the node selection phase, if a node that has not yet been fully expanded is selected, the node that has not yet been fully expanded will be expanded and a new child node will be generated;

[0036] (c) simulation;

[0037] The simulation phase starts from the expansion node and performs a downward random simulation to estimate the expected return of subsequent paths starting from the expansion node;

[0038] (d) retrospective;

[0039] In the backtracking phase, the simulation result R is transferred from the leaf node back to the root node, and the statistical reward information of the nodes on the tree path is updated.

[0040] By continuously simulating different lane-changing paths, evaluating their effects, and backtracking to update node values, the optimal lane-changing trajectory is ultimately achieved.

[0041] Furthermore, the specific method of step three is as follows:

[0042] 31) Discretize each dimension in the logical scene parameter space, and the full space parameter point set G, g∈G, g is the specific scene parameter obtained by a single discretization. The step size is determined according to the coverage accuracy requirements. The discrete step size of each dimension should not be greater than one percent of the parameter range of each dimension.

[0043] 32) Calculate each parameter point g in the discrete full space parameter point set G in turn i Is it located in a specific scene s obtained by sampling? j Within the representative range, the parameter points g are calculated in sequence. i and parameter point s j Is the distance less than the parameter point s j The radiation area calculated in formula (1) is as follows: if the parameter point g i In a specific scene s obtained by sampling j Within the representative range, it represents the parameter point g i Can be overwritten;

[0044] 33) Count all the parameter points g that can be covered in the full space parameter point set G i The number n f , and the sum of the number of all parameter points in the full space parameter point set G n all Divide them to obtain the coverage quantization result u of the sampled specific scene parameter set S in the logical scene parameter space, as shown in formula (6):

[0045]

[0046] The beneficial effects of the present invention are:

[0047] 1) This invention proposes a scenario representativeness-based test coverage evaluation method for autonomous driving systems. By setting a quantitative evaluation index for scenario representativeness, this method quantitatively evaluates the coverage of a small number of specific scenarios sampled in the logical scenario parameter space by an enterprise or a third-party testing organization.

[0048] 2) This invention can evaluate the coverage of the test process and guide companies in sampling and generating a comprehensive set of test scenario points within the parameter space, thereby establishing standardized test scenario combinations within that context. This can effectively help companies or third-party testing organizations establish standardized autonomous vehicle testing processes and promote autonomous vehicle performance verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0052] See Figure 1 The present invention provides a method for evaluating test coverage of an autonomous driving system based on scenario representativeness, comprising the following steps:

[0053] Step 1: Define the logical scene parameter space and sample the specific scene parameter set S, s∈S, s is the specific scene parameter obtained by a single sample;

[0054] Step 2: Establish a quantitative evaluation method for scene representativeness. The specific method is as follows:

[0055] 21) Without considering the differentiation of the autonomous driving system being tested, the representativeness of a specific scenario should be jointly determined by two inherent properties: probability and dangerousness. The probability affects the frequency of occurrence of the scenario in the real natural driving process, and the dangerousness indicates the accident situation of the scenario in the real driving environment. For scenarios that have both high risks and frequently appear in the natural driving database, their test results are crucial to safety verification and are difficult to be directly replaced by adjacent scenarios. Therefore, they should be given higher attention and test density, and the radiation range should be smaller during the scenario coverage calculation process, at which time their representativeness is lower. For scenarios with lower risks or lower probabilities of occurrence, redundant sampling can be reduced by a larger radiation range in the coverage strategy, and their representativeness should be stronger. Based on this rule, the present invention proposes a scenario representativeness quantification function driven jointly by risk and probability to adaptively determine the radiation range of each specific scenario in the parameter space.

[0056] The representative quantitative evaluation of the specific scenario defined is shown in formula (1):

[0057]

[0058] Where r i(j) is the radiation range of the i-th specific scene in the j-th dimension of the parameter space, that is, representativeness; i is the scene danger index of the i-th scene; A i is the scenario probability index of the i-th specific scenario; λ is the risk factor adjustment index; δ is the scaling constant of the Gaussian probability density function; ε is the probability adjustment constant; β is the global influence range control coefficient; ω j is the weight coefficient of the j-th dimension;

[0059] In formula (1), the logical scene parameter space is discretized to avoid the existence of fuzzy boundaries between adjacent points and reduce the operability and stability of the model. i (j) The influence range is rounded down to an integer to ensure that the structural boundary and computational overhead are controllable during the actual test point selection and space search process. At the same time, the global influence range minimum value r is set. min Prevent point-like impact areas due to excessively high probability or too low risk level; set the maximum value r max Prevent the extremely low probability of edge scenarios from causing the impact range to expand infinitely. (ξ i / maxξ) λ It is used to normalize the risk level parameters, making the impact scope of high-risk scenarios shrink more significantly, while avoiding the inconsistency of the impact scope across logical scenarios caused by differences in level settings. j In terms of the influence of each parameter dimension on the test results, there are significant differences. This heterogeneity between dimensions manifests as different risk gradients in the parameter space. Therefore, the present invention introduces an independent weight coefficient ω for each dimension into the adaptive function of the test point influence range. j , to adjust the expansion capabilities in different directions so that the scope of impact can more accurately match the actual importance of each test scenario.

[0060] It's important to note that the actual logical scenario parameter system includes both continuous and discrete parameters. The former possess numerical continuity, making them suitable for comprehensive test parameter analysis based on scenario representativeness. The latter, on the other hand, presents them in an enumerated form, making direct numerical comparability between dimensions impossible. To address this, we propose using discrete parameter combinations as classification labels to divide the entire logical scenario parameter space into several sub-logical scenarios. Within each sub-scenario, we independently evaluate and screen the representativeness of the test scenario, then integrate the results to avoid cross-variable variables interfering with the model structure.

[0061] 22) For the scenario probability index A in formula (1) iThe present invention uses Gaussian distribution for modeling. In real traffic behavior, the driver's response to environmental changes has certain regularity and concentration. Most operation and dynamic parameters tend to cluster around a typical value. This characteristic makes the parameter samples collected from natural driving statistically conform to the normal distribution assumption. Therefore, the present invention uses Gaussian model to effectively characterize the probability distribution of the test scenario in the logical parameter space. Scenario probability index A i The calculation method of is shown in formula (2); the relevant data can be directly obtained by statistically analyzing natural driving data.

[0062]

[0063] Where g is the parameter dimension; ∑ is the covariance matrix used to describe the correlation between various types of parameters; μ is the mean vector of each parameter variable; X is the parameter vector of the specific scenario;

[0064] 23) For the scenario risk index ξ of formula (1), the present invention considers both collision and non-collision scenarios for calculation. For collision scenarios, the present invention divides the scenarios into three categories: severe collision, general collision and minor collision according to the relative speed of the two vehicles at the time of collision. Severe collision means that the relative speed of the two vehicles at the time of collision is greater than 65km / h, general collision means 20km / h to 65km / h, and minor collision means less than 20km / h. For non-collision scenarios, the present invention calculates A based on the inverse of the time to collision (TTC). TTC -1 The calculation method is shown in formula (3):

[0065]

[0066] Where Δdis is the distance between the front and rear vehicles; Δv is the relative speed between the front and rear vehicles; 0.7s -1 TTC as a near-collision and safety scenario -1 boundary;

[0067] Taking into account both collision and non-collision situations, the scene hazard index ξ i It is defined as a five-level risk rating, with severe collision at 1.0, general collision at 0.8, minor collision at 0.6, near collision at 0.4, and safety scenario at 0.2;

[0068] As for the scene danger index, it is difficult to extract it by statistically analyzing existing data. Therefore, the present invention proposes to use a proxy model to infer the unknown scene danger index, and then obtain the danger index of different specific scenes.

[0069] The driver agent model is divided into two parts: longitudinal and lateral. The longitudinal agent model is modeled using the Long Short-Term Memory (LSTM) network, while the lateral agent model uses the lane-changing potential model and Monte Carlo Tree Search (MCTS).

[0070] For the lateral agent model, the lane change potential model quantifies the driver's satisfaction with the current driving environment and provides dynamic intention input for lane change planning; MCTS explores the optimal lane change strategy by simulating multiple possible lane change paths, namely the lateral driver agent model.

[0071] The lane-changing potential energy model is defined by comprehensively considering vehicle speed, mass, relative information of the preceding vehicle, and road adhesion conditions. The calculation formula is shown in (4), and the U threshold for generating lane-changing intention is defined as 0.15;

[0072]

[0073] Where I is a constant used to adjust the magnitude of potential energy and is set to 0.001; M i is the equivalent mass of the target vehicle, which is 5000 kg; r ij is the relative distance between the main vehicle and the target vehicle; Δv is the relative speed; θ i The angle between the speed direction of the main vehicle and the direction of the line connecting the target vehicle; l1 and l2 are attenuation factors, which are used to describe the exponential effects of distance and speed on potential energy, and are set to 1 and 0.05 respectively;

[0074] In formula (4), the environmental perception weight W is adjusted according to the road adhesion coefficient under different driving environments. Under low adhesion coefficient conditions, the driver will actively reduce the vehicle dynamic response, thereby slowing down the accumulation of potential energy; as shown in formula (5):

[0075]

[0076] Where ρ is a constant used to adjust the influence of environmental perception weight; o is the road adhesion coefficient;

[0077] After generating the lane-changing intention, MCTS is used to search for the optimal lane-changing path.

[0078] The MCTS method is divided into the following four steps:

[0079] (a) selection;

[0080] Starting from the root node, select an optimal child node along the Monte Carlo tree structure; the goal of selection is to maximize the "value" of the decision strategy;

[0081] (b) expansion;

[0082] During the node selection phase, if a node that has not been fully expanded (the node does not yet contain all possible child nodes) is selected, the node will be expanded and new child nodes will be generated;

[0083] (c) simulation;

[0084] The simulation phase starts from the expansion node and performs a downward random simulation to estimate the expected return of subsequent paths starting from this node;

[0085] (d) retrospective;

[0086] In the backtracking phase, the simulation result R is transferred from the leaf node back to the root node, and the statistical reward information of the nodes on the tree path is updated.

[0087] By continuously simulating different lane-changing paths, evaluating their effects, and backtracking to update node values, the optimal lane-changing trajectory is ultimately achieved.

[0088] Step 3: Quantitatively evaluate the test coverage. The specific method is as follows:

[0089] 31) Discretize each dimension in the logical scene parameter space, and the full space parameter point set G, g∈G, g is the specific scene parameter obtained by a single discretization. The step size is determined according to the coverage accuracy requirements. The discrete step size of each dimension should not be greater than one percent of the parameter range of each dimension.

[0090] 32) Calculate each parameter point g in the discrete full space parameter point set G in turn i Is it located in a specific scene s obtained by sampling? j Within the representative range, the parameter points g are calculated in sequence. i and parameter point s j Is the distance less than the parameter point s j The radiation area calculated in formula (1) is as follows: if the parameter point g i In a specific scene s obtained by sampling j Within the representative range, it represents the parameter point g i Can be overwritten;

[0091] 33) Count all the parameter points g that can be covered in the full space parameter point set G i The number n f , and the sum of the number of all parameter points in the full space parameter point set G n all Divide them to obtain the coverage quantization result u of the sampled specific scene parameter set S in the logical scene parameter space, as shown in formula (6):

[0092]

[0093] In summary, the present invention sets a quantitative evaluation index for scenario representativeness to quantitatively evaluate the coverage of a small number of specific scenarios sampled by enterprises or third-party testing agencies in the logical scenario parameter space. This can effectively help enterprises or third-party testing agencies establish a standardized autonomous driving vehicle testing process and promote the performance verification process of autonomous driving vehicles.

[0094] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating test coverage of an autonomous driving system based on scenario representativeness, characterized in that: The following steps are involved: Step 1: Define the logical scene parameter space and sample the specific scene parameter set S; Step 2: Establish a quantitative evaluation method for scene representativeness; Step 3: Quantitatively evaluate the test coverage.

2. The method for evaluating test coverage of an autonomous driving system based on scenario representativeness according to claim 1, characterized in that: The specific method of step 2 is as follows: 21) In the logical scenario parameter space, the representativeness of a specific scenario is determined by two inherent properties: probability and risk. Probability represents the frequency of a scenario in real driving, while risk represents the accident situation of a scenario in a real driving environment. Therefore, a scenario representativeness quantification function driven by both risk and probability is designed to adaptively determine the radiation range of each specific scenario in the parameter space, as follows: Where r i (j) is the radiation range of the i-th specific scene in the j-th dimension of the parameter space, that is, representativeness; i is the scene danger index of the i-th scene; A i is the scenario probability index of the i-th specific scenario; λ is the risk factor adjustment index; δ is the scaling constant of the Gaussian probability density function; ε is the probability adjustment constant; β is the global influence range control coefficient; ω j is the weight coefficient of the j-th dimension; In formula (1), the logical scene parameter space is discretized and r is limited. i (j) The impact range is rounded down to an integer; at the same time, the global impact range minimum value r is set min Prevent point-like impact areas; set the maximum value r max Prevent the range from expanding infinitely; set (ξ i / maxξ) λ , used to normalize the risk level parameters; introduce independent weight coefficients ω for each dimension into the adaptive function of the test point influence range to adjust the expansion capabilities in different directions; 22) For the scenario probability index A i , modeled using a Gaussian distribution as follows: Where g is the parameter dimension; ∑ is the covariance matrix used to describe the correlation between various types of parameters; μ is the mean vector of each parameter variable; X is the parameter vector of the specific scenario; 23) For the scene danger index ξ i , considering both collision and non-collision scenarios for calculation; for collision scenarios, it is divided into three categories according to the relative speed of the two vehicles at the time of collision: severe collision, general collision and minor collision. Severe collision means that the relative speed of the two vehicles at the time of collision is greater than 65km / h, general collision is 20km / h to 65km / h, and minor collision is less than 20km / h; for non-collision scenarios, A is calculated based on the inverse of the collision time, i.e., Time to Collision, TTC. TTC -1 The calculation is as follows: Where Δdis is the distance between the front and rear vehicles; Δv is the relative speed between the front and rear vehicles; 0.7s -1 TTC as a near-collision and safety scenario -1 boundary; Taking into account both collision and non-collision situations, the scene hazard index ξ i It is defined as a five-level risk rating, with severe collision at 1.0, general collision at 0.8, minor collision at 0.6, near collision at 0.4, and safety scenario at 0.2; Use the proxy model to infer the unknown scene danger index, and then obtain the danger index of different specific scenes; The driver agent model is divided into two parts: longitudinal and lateral. The longitudinal agent model is modeled using a long short-term memory network, while the lateral agent model is modeled using a lane-changing potential energy model and Monte Carlo tree search. For the lateral agent model, the lane-changing potential model quantifies the driver's satisfaction with the current driving environment and provides dynamic intention input for lane-changing planning; MCTS explores the optimal lane-changing strategy by simulating multiple possible lane-changing paths. The lane-changing potential energy model is defined by comprehensively considering vehicle speed, mass, relative information of the preceding vehicle, and road adhesion conditions. The calculation formula is shown in (4), and the U threshold for generating lane-changing intention is defined as 0.15; Where I is a constant used to adjust the magnitude of potential energy and is set to 0.001; M i is the equivalent mass of the target vehicle, which is 5000 kg; r ij is the relative distance between the main vehicle and the target vehicle; Δv is the relative speed; θ i The angle between the speed direction of the main vehicle and the direction of the line connecting the target vehicle; l1 and l2 are attenuation factors, which are used to describe the exponential effects of distance and speed on potential energy, and are set to 1 and 0.05 respectively; In formula (4), the environmental perception weight W is adjusted according to the road adhesion coefficient under different driving environments. Under low adhesion coefficient conditions, the driver will actively reduce the vehicle dynamic response, thereby slowing down the accumulation of potential energy; as shown in formula (5): Where ρ is a constant used to adjust the influence of environmental perception weight; o is the road adhesion coefficient; After generating the lane-changing intention, MCTS is used to search for the optimal lane-changing path.

3. The method for evaluating test coverage of an autonomous driving system based on scenario representativeness according to claim 2, characterized in that: The specific method of the MCTS to search for the optimal lane change path is as follows: (a) selection; Starting from the root node, select an optimal child node along the Monte Carlo tree structure; the goal of the selection is to maximize the value of the decision strategy; (b) expansion; During the node selection phase, if a node that has not yet been fully expanded is selected, the node that has not yet been fully expanded will be expanded and a new child node will be generated; (c) simulation; The simulation phase starts from the expansion node and performs a downward random simulation to estimate the expected return of subsequent paths starting from the expansion node; (d) retrospective; In the backtracking phase, the simulation result R is transferred from the leaf node back to the root node, and the statistical reward information of the nodes on the tree path is updated. By continuously simulating different lane-changing paths, evaluating their effects, and backtracking to update node values, the optimal lane-changing trajectory is ultimately achieved.

4. The method for evaluating test coverage of an autonomous driving system based on scenario representativeness according to claim 2, characterized in that: The specific method of step three is as follows: 31) Discretize each dimension in the logical scene parameter space, and the full space parameter point set G, g∈G, g is the specific scene parameter obtained by a single discretization. The step size is determined according to the coverage accuracy requirements. The discrete step size of each dimension should not be greater than one percent of the parameter range of each dimension. 32) Calculate each parameter point g in the discrete full space parameter point set G in turn i Is it located in a specific scene s obtained by sampling? j Within the representative range, the parameter points g are calculated in sequence. i and parameter point s j Is the distance less than the parameter point s j The radiation area calculated in formula (1) is as follows: if the parameter point g i In a specific scene s obtained by sampling j Within the representative range, it represents the parameter point g i Can be overwritten; 33) Count all the parameter points g that can be covered in the full space parameter point set G i The number n f , and the sum of the number of all parameter points in the full space parameter point set G n all Divide them to obtain the coverage quantization result u of the sampled specific scene parameter set S in the logical scene parameter space, as shown in formula (6):

Images