Intelligent automobile complex test scene-oriented subjective and objective fusion key quantification method

By integrating key quantization methods and soft clustering algorithms subjectively and objectively, the problems of omissions and high costs of key test scenarios in complex test scenarios of smart cars are solved, and efficient and comprehensive test scenario screening and quantification are achieved, which improves the testing of smart cars Efficiency and driver acceptance.

CN118982675AActive Publication Date: 2024-11-19JILIN UNIVERSITY

Patent Information

Application Number
CN202410980778.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-19
Estimated Expiration
2044-07-22

AI Technical Summary

Technical Problem

The existing technology mainly adopts objective and critical quantitative methods in complex test scenarios of smart cars, resulting in possible omissions of key test scenarios, reducing drivers' acceptance of smart car functions, and high testing costs and serious waste of computing resources.

Method used

A key quantitative method for subjective and objective integration of intelligent automobile complex test scenarios is proposed. By splitting the data of complex test scenarios into meteorological, traffic and road elements, combining virtual simulation software and driving simulator for subjective quantification, and using risk fields The method and interactive field strength characteristics are objectively quantified, and the efficiency and coverage are finally improved through soft clustering algorithms.

Benefits of technology

It realizes comprehensive and critical quantification of complex test scenarios, improves testing efficiency, reduces labor costs, ensures the coverage and testing value of key test scenarios, and increases drivers' acceptance of smart cars.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982675A_ABST
    Figure CN118982675A_ABST
Patent Text Reader

Abstract

The invention relates to an intelligent automobile complex test scene-oriented subjective and objective fusion criticality quantification method, which comprises the steps of processing a complex test scene data set, performing subjective and objective fusion criticality quantification, performing criticality quantification on a test scene by using a soft clustering algorithm and the like. According to the method, subjective driver evaluation and an objective risk field are fused to quantify the criticality of a complex test scene; and secondly, based on a soft clustering method, key quantification considering sample number imbalance is realized, and the labor cost is reduced while the efficiency is improved. According to the invention, the test value and necessity of the complex test scene can be determined, and the scene with high test value can be screened out. According to the invention, a research basis can be established for key test scene extraction and generation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for extracting and generating key test scenarios for smart cars, and in particular to a subjective and objective fusion criticality quantification method for complex test scenarios for smart cars. Background Art

[0002] Smart cars are an important direction for the development of the future automobile industry. Smart cars achieve smart driving tasks through perception, decision-making, planning and control systems. However, the actual driving environment includes complex weather and complex traffic environments, which will pose great challenges to the safe driving of smart cars. Therefore, before smart cars are actually put into use, it is necessary to fully test the performance of smart cars in complex environments to ensure the normal operation of smart car functions. Testing smart cars based on scenarios has become an important means. Testing smart cars using complex test scenarios can improve test efficiency and shorten the development cycle of smart cars. However, there are a large number of complex test scenarios actually collected. If the complex test scenarios collected by real cars are used to test smart cars, a lot of computing resources will be wasted, and the testing cost of complex test scenarios is high. Therefore, it is urgent to establish a criticality quantification method for complex test scenarios of smart cars. Most of the existing research methods only use objective criticality quantification methods, such as collision time, headway time, etc. However, after testing smart cars based on the key test scenarios selected by these methods, the key test scenario omission problem may still occur, which will reduce the driver's acceptance of smart car functions. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a subjective and objective fusion criticality quantification method for complex test scenarios of smart cars, comprising the following steps:

[0004] (1) Processing of complex test scenario data sets:

[0005] Split the data at time t of the nth complex test scenario into three scenario elements, including meteorological elements Traffic elements and road features

[0006] Among them, meteorological elements Including rainfall intensity R a (Unit: mm / h) and fog visibility Vis (Unit: m); Traffic elements Including the vehicle's lateral and longitudinal position (unit: m), speed (unit: m / s) and acceleration (unit: m / s 2 ); road elements Including road adhesion coefficient μ, lane line position (unit: m) and lane line linearity (including solid and dotted lines);

[0007] (2) Key quantitative methods integrating subjective and objective factors

[0008] (2.1) Quantification Method for Criticality of Subjective Complex Test Scenarios

[0009] Firstly, the above three scene elements are reproduced by using virtual simulation software, that is, a complex scene identical to the real environment is constructed in the virtual simulation scene;

[0010] The driving perspective of the host vehicle is then displayed on a driving simulator, which also has rearview mirrors and reversing mirrors, so that the person doing the subjective quantification can fully experience the behavior of the vehicles around the host vehicle.

[0011] Recruit several drivers to conduct complex scenario evaluation on the test scenario after observing the video on the driving simulator, i.e. criticality quantification; the subjective evaluation matrix focuses on two aspects of the test scenario: the degree to which the driver believes that the scenario has a collision risk and whether the driver is willing to control the vehicle to improve traffic efficiency; the subjective evaluation matrix is ​​used to determine the subjective evaluation results corresponding to the complex test scenario;

[0012] The present invention gives a greater weight to the risk dimension in the evaluation matrix. In order to ensure the rationality of subjective quantification, several different drivers are selected to quantify each test scenario, and the subjective quantification result Lab of the nth test scenario is finally obtained by weighting the number. n , the specific weighting process is as follows:

[0013]

[0014] Where j represents the scene label value; m j Select the number of drivers corresponding to label j.

[0015] Using only subjective complex test scenario criticality quantification methods will lead to a decrease in the coverage of key test scenarios. Therefore, it is necessary to use objective complex test scenario criticality quantification methods to obtain more comprehensive key test scenarios.

[0016] (2.2) Objective and complex test scenario criticality quantification method

[0017] The critical quantification method of objective complex test scenarios is constructed based on the risk field method. The risk field method is used to construct risk fields for traffic elements and road elements respectively, and the meteorological elements are used to construct the field strength attenuation coefficient; then the above field strengths are used to construct the risk fields generated by the main vehicle and the environment respectively, and the interaction between them is used to construct the interactive field strength feature. The interactive field strength feature (InteractiveFieldFeature, IFF) is defined as the quantitative result of the objective method, which integrates the three heterogeneous elements of the scene's traffic, road and meteorological elements into a quantitative result, thereby objectively describing the driving risk of the main vehicle in the current driving environment.

[0018] The calculation process of the interaction field characteristics is as follows:

[0019] First, the road potential field is generated by road elements, traffic elements and meteorological elements respectively. Traffic potential field and adverse weather field intensity degradation factor On this basis, the environmental field strength is constructed Then, follow the The same method is used to calculate the risk field generated by the main vehicle. Finally, the Delaunay triangulation discretization method is used to discretize and Two continuous inhomogeneous potential fields and calculate the interaction value IFF between two discrete potential fields n,t ; and The stronger the repulsion between n,t The larger the value, the greater the safety risk of the driving environment at the tth time point in the nth scene to the vehicle; the largest IFF value among all moments in the scene n,t As a key quantitative result of objective and complex test scenarios IFF n ; To ensure the dimensional consistency of key quantitative results, IFF n Perform normalization.

[0020] Furthermore, in step (2.2), the traffic potential field The calculation method is as follows:

[0021] Traffic potential field generated by a single vehicle m It is defined as the representation of the risk of collision with surrounding positions based on motion parameters such as position, velocity and acceleration. Its mapping relationship is expressed as

[0022]

[0023] The expression is:

[0024]

[0025] Among them, M m is the virtual mass determined by the type and actual mass of vehicle m; γ R Describes the change in the road adhesion coefficient μ caused by rain, μ r Indicates the road adhesion reduction coefficient after rainfall; c r is a constant that determines the risk growth rate when approaching m, c r Set to 0.7; is the space-time distance between any point in the space-time domain and m, and its construction process takes the vehicle speed into account.

[0026] The road potential field The calculation method is as follows:

[0027] Road potential field There are two important road elements to consider: one is the potential field generated by static obstacles. Such as obstacles; the other is the potential field generated by the lane The risk generated is evenly distributed. When the speed is set to 0, the corresponding For the lane, the potential field generated by the lane The expression is:

[0028]

[0029] Among them, N L Indicates the number of lanes; I i The meaning of the lane line linear ε of the i-th lane is that the value corresponding to the solid line is greater than the value corresponding to the dotted line; σ is a fixed value, the size of which determines the value of the vehicle when approaching the lane line growth rate; The expression is:

[0030]

[0031] Weather Conditions The intelligent vehicle perception system mainly considers two weather conditions: rain and fog. It models the impact of rain and fog on the intelligent vehicle perception module. It also models the adverse weather field intensity degradation factor caused by rain and fog weather conditions on the vehicle perception module. Affected by meteorological factors The coupling effect of the longitudinal distance Dis between the target and the vehicle is adopted. It is represented by, where the choice of W depends on the calibration of different sensors; It represents the risk aggravation factor of the scenario compared with the situation without adverse weather conditions; The modeling process is to determine Dis and The variation law of the strength degradation factor under different combinations of

[0032] Further, The construction method is applied to the YOLO-v5 target recognition function in the camera perception module of smart cars in rainy and foggy weather; in order to obtain Dis and The comprehensive parameter combination of PreScan and Simulink is used to generate images to construct W; first, different Different Dis scenarios under different conditions are analyzed, and the current Dis value and the corresponding image are recorded in combination with Simulink; secondly, the acquired image is input into YOLO-v5 to identify the target in the image, and Dis and IOU recognition results under different combinations are defined as IOU W ; Again, no record The IOU recognition results under different Dis are defined as IOU O ; Finally, Dis, and the perceptual attenuation coefficient Θ W They are represented by the x-axis, y-axis and z-axis respectively;

[0033] IOU W and IOU O The absolute difference between them is defined as Θ W , respectively, when the weather conditions are affected and not affected in the same Dis scene, Θ W The expression is:

[0034] Θ W =IOU W -IOU O (8)

[0035] Records include rainfall intensity R a and fog visibility Vis including Dis and IOU of different combinations W As a result, we construct Θ W ; The expression is:

[0036]

[0037] Among them, β T is the object category coefficient, which is used to determine the risk level of different target objects to the ego vehicle. It should be emphasized that Θ W It is a function of W and can also be calibrated by other intelligent vehicle perception systems.

[0038] According to the above three potential field calculation methods, the scene n at time t is obtained expression:

[0039]

[0040] for The establishment of its value is determined by and The maximum intensity between Sure.

[0041] Furthermore, in step (2.2), the interaction value IFF between the two discrete potential fields is n,t The calculation method is as follows:

[0042] In getting and After that, the continuous potential field is first discretized by the Delaunay triangle discretization method; then, the IFF is calculated using the characteristic points of each triangular surface. n,t ;

[0043] The Delaunay triangle has the advantages of uniqueness, optimality and proximity. and Get the corresponding triangular surface and By using Risk characteristics and Risk characteristics Quantify and IFF n,t ; and The calculation method is the same as Take the example to illustrate: the i-th The coordinates of the three vertices are and Corresponding The expression is:

[0044]

[0045] According to the law of universal gravitation, IFF n,t Depend on and Quantization, its expression is:

[0046]

[0047] in, refers to the i-th scene in the n-th scene Coordinate value at time t; refers to the jth scene in the nth scene The coordinate value at time t.

[0048] (2.3) The final scenario quantitative result C is obtained by taking the average of the subjective and objective test scenario key quantitative results n , whose expression is:

[0049]

[0050] (3) Use soft clustering algorithm to quantify the criticality of test scenarios

[0051] In order to improve the efficiency of criticality quantification, the soft clustering algorithm is used to perform criticality quantification on the test scenarios that have not been quantified. The soft clustering algorithm used takes into account the imbalance of the number of samples. The soft clustering algorithm is used to automatically quantify the criticality of other scenarios based on the quantified criticality results, and the final criticality quantification results of the complex test scenarios are output.

[0052] The soft clustering algorithm uses fuzzy set theory to expand the original membership to any value between 0 and 1, targeting the traditional hard clustering algorithm, whose membership value is either 0 or 1. A sample can belong to different clusters with different memberships, which greatly improves the clustering algorithm's processing ability for real data sets. The general equalization fuzzy C-means clustering algorithm (GEFCM) is an improvement on the optimization criterion of the traditional fuzzy C-means clustering algorithm. The sample capacity is taken into account in the design of the objective function, which can effectively alleviate the inaccurate clustering results caused by the small proportion of key scenarios in actual scene applications. The objective function set by GEFCM is:

[0053]

[0054] And satisfy the constraints:

[0055]

[0056] Where m represents the fuzzy control coefficient; c i represents the cluster center of the i-th cluster (i=1,2,…,c), c represents the total number of clusters; x j is the test scene data, j = 1, 2, ..., n represents a total of n test scene data for clustering; U = (u ij ) c*n is the membership matrix, u ij Represents the membership degree x j The degree of belonging to the i-th cluster; Dis(xj ,c i ) represents scene x j With scene c i The similarity between them needs to consider the traffic N in the scene respectively. T 、Road N R and Meteorological W The parameter values ​​are specifically defined as:

[0057]

[0058] In the formula, Represents the average value of the DTW distance between all corresponding position vehicle trajectories in the two scenes; a new objective function is constructed based on the Lagrange multiplier method according to the GEFCM objective function and constraints:

[0059]

[0060] For u in formula (16) ij and c i Taking the partial derivative and setting it to 0 gives:

[0061]

[0062] By accumulating i on both sides, we get:

[0063]

[0064] Thus we get the Lagrange multiplier λ j The expression is:

[0065]

[0066] λ j Substitute the original formula to get u ij The iteration formula is:

[0067]

[0068] The soft clustering algorithm can be used to obtain the quantitative results of the test scenarios without subjective and objective critical quantification. ij and c i The expression of is a coupled simultaneous equation and cannot be solved analytically. Therefore, u is obtained through iteration. ij and c i The estimated value of .

[0069] Beneficial effects of the present invention:

[0070] The present invention proposes a subjective and objective fusion criticality quantification method for complex test scenarios of intelligent vehicles. First, the method integrates subjective driver evaluation and objective risk field to quantify the criticality of complex test scenarios; second, based on the soft clustering method, criticality quantification considering the imbalance of sample numbers is realized, which improves efficiency while reducing labor costs. The advantages of subjective complex test scenario criticality quantification of the present invention are two-fold: on the one hand, the driver can comprehensively analyze the challenges of complex environment to driving tasks, and then screen out complex test scenarios with driving challenges; on the other hand, the test scenarios screened by the driver are used to test the intelligent vehicle. If the intelligent vehicle can fully handle the complex test scenarios screened by the driver, the driver's acceptance of the intelligent vehicle can be improved. However, only using the subjective complex test scenario criticality quantification method will lead to a decrease in the coverage of critical test scenarios. Therefore, the present invention uses the objective complex test scenario criticality quantification method to obtain a more comprehensive critical test scenario. The objective complex test scenario criticality quantification method is constructed based on the risk field method. The risk field method can fuse the three heterogeneous elements of the scene's traffic, road and meteorological elements into a quantitative result. The present invention can determine the test value and necessity of complex test scenarios, and can screen out scenarios with high test value. The present invention can establish a research foundation for the extraction and generation of key test scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 The figure is a flow chart of the subjective and objective fusion critical quantification method for complex testing scenarios of smart cars according to the present invention.

[0072] Figure 2 This is a schematic diagram of the present invention using virtual simulation software to reproduce complex test scenarios.

[0073] Figure 3 Schematic diagram of the subjective criticality quantization matrix of the present invention.

[0074] Figure 4 The figure is a schematic diagram of the interactive field feature construction process of the present invention.

[0075] Figure 5 The YOLO-v5 target recognition function performance of the present invention is different at different d W and Schematic diagram of the acquisition process under combination.

[0076] Figure 6 The present invention has different Dis and Schematic diagram of the perceptual attenuation coefficient results of YOLO-v5 under the combination.

[0077] Figure 7 For the present invention Schematic diagram of the construction method flow chart.

[0078] Figure 8 Schematic diagram of the interactive field eigenvalue calculation process between two discrete potential fields of the present invention.

[0079] Fig. 9 This is a schematic diagram of the key quantitative results obtained by the soft clustering algorithm of the present invention. DETAILED DESCRIPTION

[0080] This embodiment provides a subjective and objective fusion critical quantification method for complex test scenarios of smart cars, such as Figure 1 As shown, the following steps are included:

[0081] (1) Processing of complex test scenario data sets:

[0082] Split the data at time t of the nth complex test scenario into three scenario elements, including meteorological elements Traffic elements and road features

[0083] Among them, meteorological elements Including rainfall intensity R a (Unit: mm / h) and fog visibility Vis (Unit: m); Traffic elements Including the vehicle's lateral and longitudinal position (unit: m), speed (unit: m / s) and acceleration (unit: m / s 2 ); road elements Including road adhesion coefficient μ, lane line position (unit: m) and lane line linearity (including solid and dotted lines);

[0084] (2) Key quantitative methods integrating subjective and objective factors

[0085] (2.1) Quantification Method for Criticality of Subjective Complex Test Scenarios

[0086] Firstly, the above three scene elements are reproduced by using virtual simulation software, that is, a complex scene identical to the real environment is constructed in the virtual simulation scene;

[0087] Specific steps are as follows Figure 2 As shown in the figure, for weather and road elements, they need to be preset in advance in PreScan. For traffic elements, the FromSpreadsheet module in simulink needs to be used to read the vehicle motion data of each frame.

[0088] The driving perspective of the host vehicle is then displayed on a driving simulator, which also has rearview mirrors and reversing mirrors, so that the person doing the subjective quantification can fully experience the behavior of the vehicles around the host vehicle.

[0089] Recruit several drivers to conduct complex scenario evaluation on the test scenarios after observing the videos on the driving simulator, i.e. criticality quantification. The basis for quantification is as follows: Figure 3 As shown, the subjective evaluation matrix focuses on two aspects of the test scenario: the degree to which the driver believes that the scenario has a collision risk and whether the driver is willing to control the vehicle to improve traffic efficiency; the subjective evaluation matrix is ​​used to determine the subjective evaluation results corresponding to the complex test scenario;

[0090] The present invention assigns a greater weight to the risk dimension in the evaluation matrix. In order to ensure the rationality of subjective quantification, this embodiment selects 8 different drivers to quantify each test scenario, and finally obtains the subjective quantification result Lab of the nth test scenario by weighting the number of drivers. n , the specific weighting process is as follows:

[0091]

[0092] Wherein, j represents the scene label value, and this embodiment sets 4 labels; m j Select the number of drivers corresponding to label j.

[0093] Using only subjective complex test scenario criticality quantification methods will lead to a decrease in the coverage of key test scenarios. Therefore, it is necessary to use objective complex test scenario criticality quantification methods to obtain more comprehensive key test scenarios.

[0094] (2.2) Objective and complex test scenario criticality quantification method

[0095] The critical quantification method of objective complex test scenarios is constructed based on the risk field method. The risk field method is used to construct risk fields for traffic elements and road elements respectively, and the meteorological elements are used to construct the field strength attenuation coefficient; then the above field strengths are used to construct the risk fields generated by the main vehicle and the environment respectively, and the interaction between them is used to construct the interactive field strength feature. The interactive field strength feature (InteractiveFieldFeature, IFF) is defined as the quantitative result of the objective method, which integrates the three heterogeneous elements of the scene's traffic, road and meteorological elements into a quantitative result, thereby objectively describing the driving risk of the main vehicle in the current driving environment.

[0096] like Figure 4 As shown, the calculation process of the interaction field characteristics is as follows:

[0097] First, the road potential field is generated by road elements, traffic elements and meteorological elements respectively. Traffic potential field and adverse weather field intensity degradation factor On this basis, the environmental field strength is constructed Then, follow the The same method is used to calculate the risk field generated by the main vehicle. Finally, the Delaunay triangulation discretization method is used to discretize and Two continuous inhomogeneous potential fields and calculate the interaction value IFF between two discrete potential fields n,t ; and The stronger the repulsion between n,t The larger the value, the greater the safety risk posed to the vehicle by the driving environment at the tth time point in the nth scenario.

[0098] The traffic potential field The calculation method is as follows:

[0099] Traffic potential field generated by a single vehicle m It is defined as the representation of the risk of collision with surrounding positions based on motion parameters such as position, velocity and acceleration. Its mapping relationship is expressed as

[0100]

[0101] The expression is:

[0102]

[0103] Among them, M m is the virtual mass determined by the type and actual mass of vehicle m; γ R Describes the change in the road adhesion coefficient μ caused by rain, μ r Indicates the road adhesion reduction coefficient after rainfall; c r is a constant that determines the risk growth rate when approaching m, c r Set to 0.7; is the space-time distance between any point in the space-time domain and m, and its construction process takes the vehicle speed into account.

[0104] The road potential field The calculation method is as follows:

[0105] Road potential field There are two important road elements to consider: one is the potential field generated by static obstacles. Such as obstacles; the other is the potential field generated by the lane The risk generated is evenly distributed. When the speed is set to 0, the corresponding For lanes, different lane line linear ε will cause different degrees of risk to the main vehicle. For example, when the main vehicle approaches the dotted line and prepares to change lanes, the risk that needs to be overcome is much smaller than when approaching the double solid line. This is because driving a vehicle across the double solid line will violate road traffic regulations, resulting in greater driving risks. The expression is:

[0106]

[0107] Among them, N L Indicates the number of lanes; I i The meaning of the lane line linear ε of the i-th lane is that the value corresponding to the solid line is greater than the value corresponding to the dotted line; σ is a fixed value, the size of which determines the value of the vehicle when approaching the lane line growth rate; The expression is:

[0108]

[0109] Weather Conditions The intelligent vehicle perception system mainly considers two weather conditions: rain and fog. It models the impact of rain and fog on the intelligent vehicle perception module. It also models the adverse weather field intensity degradation factor caused by rain and fog weather conditions on the vehicle perception module. Affected by meteorological factors The coupling effect of the longitudinal distance Dis between the target and the vehicle is adopted. It is represented by, where the choice of W depends on the calibration of different sensors; It represents the risk aggravation factor of the scenario compared with the situation without adverse weather conditions; The modeling process is to determine Dis and The variation law of the strength degradation factor under different combinations of

[0110] This embodiment will The construction method is applied to the YOLO-v5 target recognition function in the camera perception module of smart cars in rainy and foggy weather; in order to obtain Dis and The comprehensive parameter combination uses PreScan and Simulink to generate images to construct W, such as Figure 5 First, construct different Different Dis scenarios under different conditions are analyzed, and the current Dis value and the corresponding image are recorded in combination with Simulink; secondly, the acquired image is input into YOLO-v5 to identify the target in the image, and Dis and IOU recognition results under different combinations are defined as IOU W ; Again, no record The IOU recognition results under different Dis are defined as IOUO ; Finally, Dis, and the perceptual attenuation coefficient Θ W Represented by the x-axis, y-axis and z-axis respectively, such as Figure 6 shown.

[0111] IOU W and IOU O The absolute difference between them is defined as Θ W , respectively, when the weather conditions are affected and not affected in the same Dis scene, Θ W The expression is:

[0112] Θ W =IOU W -IOU O (8)

[0113] Records include rainfall intensity R a and fog visibility Vis including Dis and IOU of different combinations W As a result, we construct Θ W ; The expression is:

[0114]

[0115] Among them, β T is the object category coefficient, which is used to determine the risk level caused by different target objects to the ego vehicle.

[0116] According to the above three potential field calculation methods, the scene n at time t is obtained expression:

[0117]

[0118] for The establishment of its value is determined by and The maximum intensity between OK. Figure 7 As shown, different Including R a =15mm / h and Vis=30m The specific results show that The original strength of Enlarge, while With Dis and This method can accurately describe the potential field strength distribution in the environment.

[0119] In getting and After that, the continuous potential field is first discretized by the Delaunay triangle discretization method; then, the IFF is calculated using the characteristic points of each triangular surface. n,t .

[0120] The Delaunay triangle has the advantages of uniqueness, optimality and proximity. and Get the corresponding triangular surface and By using Risk characteristics and Risk characteristics Quantify and IFF n,t ,like Figure 8 As shown; and The calculation method is the same as Take the example to illustrate: the i-th The coordinates of the three vertices are and Corresponding The expression is:

[0121]

[0122]

[0123] According to the law of universal gravitation, IFF n,t Depend on and Quantization, its expression is:

[0124]

[0125] in, refers to the i-th scene in the n-th scene Coordinate value at time t; refers to the jth scene in the nth scene Coordinate value at time t;

[0126] Take the largest IFF among all moments in the scene n,t As a key quantitative result of objective and complex test scenarios IFF n ; To ensure the dimensional consistency of key quantitative results, IFF n Perform normalization.

[0127] (2.3) The final scenario quantitative result C is obtained by taking the average of the subjective and objective test scenario key quantitative resultsn , whose expression is:

[0128]

[0129] (3) Use soft clustering algorithm to quantify the criticality of test scenarios

[0130] In the actual process of quantifying the criticality of complex test scenarios, it was found that the subjective method is time-consuming and labor-intensive, and recruiting different drivers to quantify the test scenarios will incur a lot of manpower costs. Therefore, in order to improve the efficiency of criticality quantification, it is necessary to use a soft clustering algorithm to quantify the criticality of the test scenarios that have not been quantified. The soft clustering algorithm used in the present invention takes into account the imbalance of the number of samples. The reason is that the probability of occurrence of critical test scenarios in the actual collected scenarios is extremely low. If only traditional clustering algorithms are used, key features will be ignored. The soft clustering algorithm is used to automatically quantify the criticality of other scenarios based on the quantified criticality results, and the final complex test scenario criticality quantification results are output.

[0131] The soft clustering algorithm uses fuzzy set theory to expand the original membership to any value between 0 and 1, targeting the traditional hard clustering algorithm, whose membership value is either 0 or 1. A sample can belong to different clusters with different memberships, which greatly improves the clustering algorithm's processing ability for real data sets. The general equalization fuzzy C-means clustering algorithm (GEFCM) is an improvement on the optimization criterion of the traditional fuzzy C-means clustering algorithm. The sample capacity is taken into account in the design of the objective function, which can effectively alleviate the inaccurate clustering results caused by the small proportion of key scenarios in actual scene applications. The objective function set by GEFCM is:

[0132]

[0133] And satisfy the constraints:

[0134]

[0135] Where m represents the fuzzy control coefficient; c i represents the cluster center of the i-th cluster (i=1,2,…,c), c represents the total number of clusters; x j is the test scene data, j = 1, 2, ..., n represents a total of n test scene data for clustering; U = (u ij ) c*n is the membership matrix, u ij Represents the membership degree x j The degree of belonging to the i-th cluster; Dis(xj ,c i ) represents scene x j With scene c i The similarity between them needs to consider the traffic N in the scene respectively. T 、Road N R and Meteorological W The parameter values ​​are specifically defined as:

[0136]

[0137] In the formula, Represents the average value of the DTW distance between all corresponding position vehicle trajectories in the two scenes; a new objective function is constructed based on the Lagrange multiplier method according to the GEFCM objective function and constraints:

[0138]

[0139] For u in formula (16) ij and c i Taking the partial derivative and setting it to 0 gives:

[0140]

[0141] By accumulating i on both sides, we get:

[0142]

[0143] Thus we get the Lagrange multiplier λ j The expression is:

[0144]

[0145] λ j Substitute the original formula to get u ij The iteration formula is:

[0146]

[0147] The soft clustering algorithm can be used to obtain quantitative results of test scenarios that have not been quantified for subjective and objective criticality, such as Fig. 9 As shown. It can be seen that u ij and c i The expression of is a coupled simultaneous equation and cannot be solved analytically. Therefore, u is obtained through iteration. ij and c i The specific iterative solution process is shown in the following table:

[0148]

[0149]

Claims

1. A subjective and objective fusion criticality quantification method for complex test scenarios of intelligent vehicles, characterized by: The following steps are involved: (1) Processing of complex test scenario data sets: Split the data at time t of the nth complex test scenario into three scenario elements, including meteorological elements Traffic elements and road features (2) Key quantitative methods integrating subjective and objective factors (2.1) Quantification Method for Criticality of Subjective Complex Test Scenarios Firstly, the above three scene elements are reproduced by using virtual simulation software, that is, a complex scene identical to the real environment is constructed in the virtual simulation scene; The driving perspective of the main vehicle is then displayed on the driving simulator; Recruit several drivers to conduct complex scenario evaluation on the test scenario after observing the video on the driving simulator, i.e. criticality quantification; the subjective evaluation matrix focuses on two aspects of the test scenario: the degree to which the driver believes that the scenario has a collision risk and whether the driver is willing to control the vehicle to improve traffic efficiency; the subjective evaluation matrix is ​​used to determine the subjective evaluation results corresponding to the complex test scenario; In order to ensure the rationality of subjective quantification, several different drivers are selected to quantify each test scenario, and the subjective quantification result Lab of the nth test scenario is finally obtained by weighting the number of drivers. n , the specific weighting process is as follows: Where j represents the scene label value; m j Select the number of drivers corresponding to label j; (2.2) Objective and complex test scenario criticality quantification method The critical quantitative method of objective complex test scenarios is constructed based on the risk field method. The risk field method is used to construct risk fields for traffic elements and road elements respectively, and the field strength attenuation coefficient is constructed using meteorological elements. Then, the above field strengths are used to construct the risk fields generated by the main vehicle and the environment respectively, and the interaction between them is used to construct the interactive field strength feature. The interactive field strength feature IFF is defined as the quantitative result of the objective method, which integrates the three heterogeneous elements of the scene's traffic, road and meteorological elements into a quantitative result, thereby objectively describing the driving risk of the main vehicle in the current driving environment. The calculation process of the interaction field characteristics is as follows: First, the road potential field is generated by road elements, traffic elements and meteorological elements respectively. Traffic potential field and adverse weather field intensity degradation factor On this basis, the environmental field strength is constructed Then, follow the The same method is used to calculate the risk field generated by the main vehicle. Finally, the Delaunay triangulation discretization method is used to discretize and Two continuous inhomogeneous potential fields and calculate the interaction value IFF between two discrete potential fields n,t ; and The stronger the repulsion between n,t The larger the value, the greater the safety risk posed to the vehicle by the driving environment at the tth time point in the nth scenario; Take the largest IFF among all moments in the scene n,t As a key quantitative result of objective and complex test scenarios IFF n ; To ensure the dimensional consistency of key quantitative results, IFF n Perform normalization processing; (2.3) The final scenario quantitative result C is obtained by taking the average of the subjective and objective test scenario key quantitative results n , whose expression is: (3) Use soft clustering algorithm to quantify the criticality of test scenarios With the help of soft clustering algorithm, criticality quantification is performed on the test scenarios that have not been quantified. The soft clustering algorithm used takes into account the imbalance of the number of samples. The soft clustering algorithm is used to automatically quantify the criticality of other scenarios based on the quantified criticality results, and the final criticality quantification results of the complex test scenarios are output.

2. According to claim 1, a subjective and objective fusion criticality quantification method for complex test scenarios of intelligent vehicles is characterized by: In step (1), meteorological elements Including rainfall intensity R a and fog visibility Vis; traffic elements Including the vehicle's lateral and longitudinal position, speed and acceleration; road elements Including road adhesion coefficient μ, lane line position and lane line linearity.

3. According to claim 1, a subjective and objective fusion criticality quantification method for complex test scenarios of intelligent vehicles is characterized by: In step (2.2), the traffic potential field The calculation method is as follows: Traffic potential field generated by a single vehicle m It is defined as the representation of the risk of collision with surrounding positions based on motion parameters such as position, velocity and acceleration. Its mapping relationship is expressed as The expression is: Among them, M m is the virtual mass determined by the type and actual mass of vehicle m; γ R Describes the change in the road adhesion coefficient μ caused by rain, μ r Indicates the road adhesion reduction coefficient after rainfall; c r is a constant that determines the risk growth rate when approaching m, c r Set to 0.7; is the space-time distance between any point in the space-time domain and m, and its construction process takes into account the vehicle speed; The road potential field The calculation method is as follows: Road potential field There are two important road elements to consider: one is the potential field generated by static obstacles. Such as obstacles; the other is the potential field generated by the lane The risk generated is evenly distributed. When the speed is set to 0, the corresponding For the lane, the potential field generated by the lane The expression is: Among them, N L Indicates the number of lanes; I i The meaning of the lane line linear ε of the i-th lane is that the value corresponding to the solid line is greater than the value corresponding to the dotted line; σ is a fixed value, the size of which determines the value of the vehicle when approaching the lane line growth rate; The expression is: Weather Conditions The intelligent vehicle perception system mainly considers two weather conditions: rain and fog. It models the impact of rain and fog on the intelligent vehicle perception module. It also models the adverse weather field intensity degradation factor caused by rain and fog weather conditions on the vehicle perception module. by The coupling effect of the longitudinal distance Dis between the target and the vehicle is adopted. It is represented by, where the choice of W( ) depends on the calibration of different sensors; It represents the risk aggravation factor of the scenario compared with the situation without adverse weather conditions; The modeling process is to determine Dis and The variation law of the strength degradation factor under different combinations of According to the above three potential field calculation methods, the scene n at time t is obtained expression: for The establishment of its value is determined by and The maximum intensity between Sure.

4. According to claim 3, a subjective and objective fusion criticality quantification method for complex test scenarios of intelligent vehicles is characterized by: Will The construction method is applied to the YOLO-v5 target recognition function in the camera perception module of smart cars in rainy and foggy weather; in order to obtain Dis and The comprehensive parameter combination is constructed by using PreScan and Simulink to generate images to construct W( ); first, construct different Different Dis scenarios under different conditions are analyzed, and the current Dis value and the corresponding image are recorded in combination with Simulink; secondly, the acquired image is input into YOLO-v5 to identify the target in the image, and Dis and IOU recognition results under different combinations are defined as IOU W ; Again, no record The IOU recognition results under different Dis are defined as IOU O ; Finally, Dis, and the perceptual attenuation coefficient Θ W They are represented by the x-axis, y-axis and z-axis respectively; IOU W and IOU O The absolute difference between them is defined as Θ W , respectively, when the weather conditions are affected and not affected in the same Dis scene, Θ W The expression is: I W =|IOU W -IOU O | Records include rainfall intensity R a and fog visibility Vis including Dis and IOU of different combinations W As a result, we construct Θ W ; The expression is: Among them, β T is the object category coefficient, which is used to determine the risk level caused by different target objects to the ego vehicle.

5. According to claim 1, a subjective and objective fusion criticality quantification method for complex test scenarios of intelligent vehicles is characterized by: In step (2.2), the interaction value IFF between the two discrete potential fields is n,t The calculation method is as follows: In getting and After that, the continuous potential field is first discretized by the Delaunay triangle discretization method; then, the IFF is calculated using the characteristic points of each triangular surface. n,t ; The Delaunay triangle has the advantages of uniqueness, optimality and proximity. and Get the corresponding triangular surface and By using Risk characteristics and Risk characteristics Quantify and IFF n,t ; and The calculation method is the same as Take the example to illustrate: the i-th The coordinates of the three vertices are and Corresponding The expression is: According to the law of universal gravitation, IFF n,t Depend on and Quantization, its expression is: in, refers to the i-th scene in the n-th scene Coordinate value at time t; refers to the jth scene in the nth scene The coordinate value at time t.

6. The subjective and objective fusion criticality quantification method for complex test scenarios of intelligent vehicles according to claim 1 is characterized by: In step (3), the soft clustering algorithm uses fuzzy set theory to expand the original membership value to any value between 0 and 1 for the traditional hard clustering algorithm with a strict membership relationship of either 0 or 1, thereby improving the processing capability of the clustering algorithm for real data sets; The general equilibrium fuzzy C-means clustering algorithm GEFCM improves the optimization criteria of the traditional fuzzy C-means clustering algorithm. The objective function set by GEFCM is: And satisfy the constraints: Where m represents the fuzzy control coefficient; c i represents the cluster center of the i-th cluster, j = 1, 2, ..., c, c represents the total number of clusters; x j is the test scene data, j = 1, 2, ..., n represents a total of n test scene data for clustering; U = (u ij ) c*n is the membership matrix, u ij Represents the membership degree x j The degree of belonging to the i-th cluster; Dis(x j ,c i ) represents scene x j With scene c i The similarity between them needs to consider the traffic N in the scene respectively. T 、Road N R and Meteorological W The parameter values ​​are specifically defined as: In the formula, Represents the average value of the DTW distance between all corresponding position vehicle trajectories in the two scenes; a new objective function is constructed based on the Lagrange multiplier method according to the GEFCM objective function and constraints: For the GEFCM objective function formula, u ij and c i Taking the partial derivative and setting it to 0 gives: By accumulating i on both sides, we get: Thus we get the Lagrange multiplier λ j The expression is: λ j Substitute the original formula to get u ij The iteration formula is: The soft clustering algorithm can be used to obtain the quantitative results of the test scenarios without subjective and objective critical quantification. ij and c i The expression of is a coupled simultaneous equation and cannot be solved analytically. Therefore, u is obtained through iteration. ij and c i The estimated value of .

Citation Information

Patent Citations

  • Construction method of automatic driving automobile test scene library

    CN114818381A

  • Man-machine interaction test method for self-driving automobile

    CN116680179A

  • Method for evaluating performance of self-driving vehicle oriented to full parameter space of logical scenario

    US20230304896A1

  • Method and a system for testing a driver assistance system for a vehicle

    US20230394896A1

Cited By

  • Vehicle testing method, electronic equipment and vehicle

    CN122524459A

  • Vehicle testing method, electronic device and vehicle

    CN122524459B