Extreme scene search method based on improved genetic algorithm

Abstract

The present invention relates to an extreme scene search method based on an improved genetic algorithm, and belongs to the field of autonomous driving test evaluation. First, the corresponding scene parameters are set according to the corresponding driving scene, and the corresponding number of parameters and the number of scenes are set; each scene is randomly paired as a separate population, and a full gene crossover is performed to obtain a crossover offspring population; the crossover offspring population is screened to obtain a crossover preferred offspring population; the preferred offspring population is subjected to multiple mutation operations to obtain multiple groups of mutant offspring populations; multiple groups of mutant offspring populations are screened to obtain a mutant preferred offspring population; the mutant preferred offspring population is used to perform autonomous driving tests, and the current most extreme scene is selected according to the elite retention strategy; whether the convergence condition is met is determined according to the test results, and if so, the search ends, otherwise the full gene crossover process is returned to repeat the execution. The present invention effectively avoids falling into a local optimum while improving the rapidity of the extreme scene search of the autonomous driving system.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving test evaluation and relates to an extreme scene search method based on an improved gene genetic algorithm. Background Art

[0002] In the test and evaluation of autonomous driving systems, it is of great significance to identify extreme scenarios that may cause system failure or serious performance degradation. By feeding back extreme scenarios to users and developers, on the one hand, it can prevent users from over-reliance on autonomous driving systems and causing traffic accidents, and on the other hand, it can help system developers locate the deficiencies of existing systems and point out the direction of improvement for future research and development work. However, compared with traditional automotive electronic systems, the performance and functional performance of autonomous driving systems are affected by dynamic traffic scenarios. The uncontrollability (such as the inability to control the motion state of other vehicles outside the vehicle) and variability (such as the vehicle motion state, terrain, and weather) of traffic scenarios make it difficult to accurately define the operating conditions of the autonomous driving system, and it is even more difficult to list all its operating conditions one by one, making it extremely difficult to discover and find extreme scenarios.

[0003] Among the common test evaluation methods, the efficiency of road vehicle testing is extremely low. This is because most of the real scenes are conventional working conditions, and the frequency of dangerous working conditions is very low. Relatively speaking, compared with road testing, scenario-based testing methods can significantly improve test efficiency by reasonably designing and optimizing test scenarios, and more quickly discover extreme scenarios that lead to unacceptable behavior of the autonomous driving system. However, typical methods such as combination testing, radical behavior acceleration, and evolutionary testing are all based on modern intelligent optimization algorithms, and cannot simultaneously improve exploration capabilities and convergence performance. And because the autonomous driving system contains complex calculations such as deep convolutional neural networks, the impact of the scene on its performance is very complex, and there are serious non-convexity, nonlinearity, and singularity, which will inevitably lead to the search process for extreme scenarios. Convergence is very slow, and even potential extreme scenarios cannot be found. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide an extreme scenario search method based on an improved genetic algorithm, to measure the evaluation index of the effectiveness of the offspring without performing tests, and to use the evaluation index to design full crossover and multiple mutation operators to improve the exploration ability and convergence of the genetic algorithm, thereby improving the rapidity of the extreme scenario search of the autonomous driving system while effectively avoiding falling into the local optimum.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] An extreme scene search method based on an improved genetic algorithm comprises the following steps:

[0007] S1. Set the number of scene parameters N, convergence conditions, and the number of multiple mutation populations N p , a scene parameter value set, and then randomly generate 2N test scenes within the scene parameter value set;

[0008] S2, treat each scene as a separate population, pair them up randomly, and perform full gene crossover to obtain the crossover offspring population;

[0009] S3, performing a first screening in the crossover offspring population by using the scene complexity of the population to obtain a crossover optimal offspring population;

[0010] S4, performing multiple mutation operations on the selected preferred progeny populations to obtain multiple groups of mutant progeny populations;

[0011] S5. Perform a second screening among multiple groups of mutant offspring populations based on the scene complexity of the population to obtain a mutant optimal offspring population;

[0012] S6. Use the mutant-optimized offspring population to generate test scenarios to test the autonomous driving system, and select the most extreme scenario at the moment based on the elite retention strategy; the elite retention strategy means that the discovered optimal solution is retained in each iteration.

[0013] S7: Determine whether the convergence condition is met based on the test results. If so, the search ends; otherwise, return to S2.

[0014] Further, in step S2, 2N scenes are randomly paired in pairs, and for each pair of scenes, all N scenes that can be cross-operated are c Single-point crossover mutation is performed at each gene position to obtain N×N c A set of alternative offspring, in which single-point crossover mutation refers to exchanging all genes on the right side of the two parental chromosomes for the selected gene crossover point.

[0015] Further, in step S3, in N×N c Calculate the complexity of the scenes in each group of candidate children, and get the overall complexity of each group of children by summing them up. Then, perform the first screening from high to low according to the overall complexity, and select N groups of scenes. The calculation method of the overall complexity D(X) is:

[0016]

[0017] In the formula, the value of each factor x i Construct a test scenario X = {x1, ..., x m}, To construct the discrete value points of scene influencing factors using the hierarchical analysis method, for The importance of The value of is calculated through expert scoring method or entropy weight method.

[0018] Further, in step S4, N operations are performed on the 2N scenes screened out for the first time. p The basic bit mutation operation is performed to obtain N p A mutation population; wherein the basic bit mutation operation refers to performing a mutation operation on a certain bit or several bits of genes randomly specified in the individual coding string.

[0019] Further, in step S5, N p The mutant population is screened for the second time, that is, the complexity of each scene in the mutant population is calculated, the overall complexity of the population is obtained by summing up, and then the population with the highest overall complexity is selected as the mutant offspring. The overall complexity calculation method of the second screening is the same as that of the first screening process.

[0020] The beneficial effects of the present invention are:

[0021] The present invention proposes an evaluation index for measuring the effectiveness of offspring without performing tests, and uses the evaluation index to design full crossover and multiple mutation operators based on a genetic algorithm to improve the overall effectiveness of the offspring, improve the exploration ability and convergence of the genetic algorithm, thereby improving the rapidity of searching for extreme scenarios in the autonomous driving system and effectively avoiding falling into local optimality.

[0022] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0024] Figure 1 This is a flow chart of the extreme scene search method based on the improved genetic algorithm;

[0025] Figure 2 This is a schematic diagram of a parallel parking scenario;

[0026] Figure 3 Parallel parking simulation test system;

[0027] Figure 4 It is the statistical result of scene complexity and minimum distance;

[0028] Figure 5 is the relationship diagram between evolutionary generations and minimum distance;

[0029] Figure 6 The relationship between lane recognition error rate and overall scene complexity;

[0030] Figure 7 Comparison of the distribution of scenario complexity of different search algorithms. DETAILED DESCRIPTION

[0031] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0032] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0033] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0034] See also Figure 1 to Figure 7 , which is an extreme scene search method based on an improved genetic algorithm.

[0035] like Figure 1 As shown, the extreme scene search method based on the improved genetic algorithm of the present invention comprises the following steps:

[0036] S1. Set the number of scene parameters N, convergence conditions, and the number of multiple mutation populations N p , a scene parameter value set, and then randomly generate 2N test scenes within the scene parameter value set;

[0037] S2, treat each scene as a separate population, pair them up randomly, and perform full gene crossover to obtain the crossover offspring population;

[0038] S3, performing a first screening in the crossover offspring population by using the scene complexity of the population to obtain a crossover optimal offspring population;

[0039] S4, performing multiple mutation operations on the selected preferred progeny populations to obtain multiple groups of mutant progeny populations;

[0040] S5. Perform a second screening among multiple groups of mutant offspring populations based on the scene complexity of the population to obtain a mutant optimal offspring population;

[0041] S6. Use the mutant-optimized offspring population to generate test scenarios to test the autonomous driving system, and select the most extreme scenario at the moment based on the elite retention strategy; the elite retention strategy means that the discovered optimal solution is retained in each iteration.

[0042] S7: Determine whether the convergence condition is met based on the test results. If so, the search ends; otherwise, return to S2.

[0043] The extreme scenario search method based on the improved genetic algorithm of the present invention can search for extreme scenarios of different driving scenarios or different test scenarios, measure the evaluation index of the effectiveness of the offspring without performing the test, and use the evaluation index to design full crossover and multiple mutation operators based on the genetic algorithm to improve the overall effectiveness of the offspring, improve the exploration ability and convergence of the genetic algorithm, thereby improving the rapidity of the extreme scenario search of the autonomous driving system while effectively avoiding falling into the local optimum.

[0044] Example 1

[0045] This embodiment takes the parallel parking system as the object, takes the shortest distance between the vehicle and the curb during parking as the indicator for evaluating the extremeness of the scene, and provides an extreme scene search method based on an improved genetic algorithm. Figure 2 As shown, the parameters describing the scene include: the heading angle of the vehicle F2, the heading angle of the target vehicle F1, the distance from the target vehicle to the curb F3, the initial distance from the vehicle to the target vehicle F4, and the parking speed F5.

[0046] Specifically, in step S1, the number of scenes is set to N = 50, the number of multiple mutation populations is set to N p =5; randomly generate 100 test scenarios in the scenario parameter set. Set the convergence condition as follows: the evolutionary generation exceeds 25 or the closest distance between the vehicle and the curb during parking is less than or equal to 0.

[0047] In step S2, this embodiment randomly pairs the above 100 scenes into 50 groups, describes the scene parameter values ​​in binary mode for each group of scenes, and then performs single-point crossover mutation at each binary position to obtain candidate offspring.

[0048] In step S3, the obtained candidate offspring are screened for the first time, and the complexity of each scene in each group of candidate offspring is calculated. The overall complexity of each group of offspring is obtained by summing up, and then 50 groups of scenes are selected from high to low in terms of overall complexity as crossover offspring for mutation operation. The calculation formula of the complexity D(X) of each scene is:

[0049]

[0050] In the formula, the value of each factor x i Construct a test scenario X = {x1, ..., x m}, To construct the discrete value points of scene influencing factors using the hierarchical analysis method, for The importance of The value of is calculated by expert scoring method or entropy weight method. The importance of the above scenario parameters obtained by using the hierarchical analysis method and expert scoring method is shown in Table 1:

[0051] Table 1

[0052]

[0053] In Table 1, the first level refers to the primary elements used to describe the parking scene, R 1，j Refers to the relative importance of the first-level elements of the parking scene

[0054] The second level refers to the secondary elements used to describe the parking scene, R 2，j It refers to the relative importance of the secondary elements of the parking scene. The discrete value refers to the discrete value point of the secondary elements of the parking scene. j Refers to the relative importance of the discretized value points. The importance is calculated according to Table 1 The formula is:

[0055]

[0056] In the formula, Represents the discretization points A set of node indices on the path to the first-level element.

[0057] In step S4, multiple mutation operations are performed on the 50 groups of scenes screened out for the first time. Specifically, 100 scenes are obtained from S3, and the scene parameter values ​​are described in binary. Then, five basic bit mutation operations are randomly performed on the binary bits to obtain five mutant populations.

[0058] In step S5, the obtained 5 variant populations, that is, 500 scenes, are screened for the second time in the same way as the first screening, and finally 100 test scenes are obtained.

[0059] In step S6, the 100 test scenarios obtained in S5 are used to test the autonomous driving system. The most extreme scenario is selected based on the system performance using the elite retention strategy. The test scenario is simulated by Prescan software. The parallel parking algorithm and vehicle model are run in the Matlab / Simulink environment. The Simulink interface provided by Prescan is used to realize information interaction. The simulation system is as follows: Figure 3 shown.

[0060] In step S7, it is determined whether the convergence condition is met based on the test result of S6: if the evolutionary generation exceeds 25 or the closest distance between the vehicle and the curb during parking is less than or equal to 0, the search ends; otherwise, return to S2.

[0061] After the extreme scene search converges, the complexity of all scenes and the minimum distance from the road edge are statistically analyzed as follows: Figure 4 As shown in Figure 2, it can be seen that the scene complexity index can reflect the extremeness of the scene. The comparison results with the traditional genetic inheritance are as follows: Figure 5 As shown, it can be seen that the present invention measures the effectiveness of the offspring without performing tests by proposing a scene complexity index, and designs full crossover and multiple mutation operators based on the evaluation index, which effectively improves the overall effectiveness of the offspring, improves the exploration ability and convergence of the genetic algorithm, and effectively avoids falling into the local optimum while improving the rapidity of extreme scene search of the parallel automatic parking system.

[0062] Example

[0063] This embodiment takes lane line recognition scenarios as the object, takes lane line recognition error rate as the indicator for evaluating the extremeness of scenarios, and provides an extreme scenario search method based on an improved genetic algorithm. The elements describing the lane line recognition algorithm test scenario and the importance of each element obtained by the hierarchical analysis method and the entropy weight method are shown in Table 2.

[0064] Table 2

[0065]

[0066] As shown in Table 2, in this embodiment, for the lane line recognition scene test, the scene parameters given in this embodiment include at least the traffic environment and the self-vehicle behavior, wherein the traffic environment includes at least the illumination and lane line parameters, and the illumination is further divided into at least sunny, cloudy, rainy and other weather conditions according to the climate conditions, and the specific settings can be adjusted according to the actual situation; the lane line parameters are divided according to the degree of completeness, color, and type, wherein the value of the degree of completeness is a single-sided line, the value of the color is white, yellow, etc., and the value of the type is dot-dash line, solid line, etc., and the specific value can be adjusted according to the actual situation. The self-vehicle behavior parameters mainly include the longitudinal speed and the lateral speed, etc., and the specific values ​​are adjusted according to the actual situation.

[0067] Specifically, in step S1, the number of scenarios N=100 and the number of multiple mutation populations Np=10 are set; 200 test scenarios are randomly generated within the scenario parameter set. The convergence condition is set as follows: the number of evolution generations exceeds 25 or the error of the lane line recognition error rate of two consecutive tests is less than 0.0001.

[0068] In step S2, this embodiment randomly pairs the above 200 scenes and divides them into 10 groups. For each group of scenes, the scene parameter values ​​are described in binary mode, and then single-point crossover mutation is performed at each binary position to obtain candidate offspring.

[0069] In step S3, the obtained candidate offspring are screened for the first time, and the complexity of each scene in each group of candidate offspring is calculated. The overall complexity of each group of offspring is obtained by summing up, and then 100 groups of scenes are selected from high to low according to the overall complexity as crossover offspring for mutation operation. The calculation formula of the complexity D(X) of each scene is:

[0070]

[0071] In the formula, the value of each factor x i Construct a test scenario X = {x1, ..., x m}, To construct the discrete value points of scene influencing factors using the hierarchical analysis method, for The importance of The value of is calculated by expert scoring method or entropy weight method. The importance of the above scenario parameters obtained by analytic hierarchy process and expert scoring method is shown in Table 2.

[0072] In step S4, multiple mutation operations are performed on the 100 groups of scenes screened out for the first time. Specifically, 200 scenes are obtained from S3, and the scene parameter values ​​are described in binary. Then, 10 basic bit mutation operations are randomly performed on the binary bits to obtain 10 mutant populations.

[0073] In step S5, the obtained 10 variant populations, that is, 2000 scenes, are screened for the second time in the same way as the first screening, and finally 200 test scenes are obtained.

[0074] In step S6, the lane recognition algorithm is tested using the 200 test scenarios obtained in S5, and the current most extreme scenario set is selected using the elite retention strategy based on the system performance. The test scenarios are simulated by Prescan software, and the lane recognition algorithm runs in the Matlab / Simulink environment, and information interaction is achieved through the Simulink interface provided by Prescan.

[0075] In step S7, it is determined whether the convergence condition is met based on the test result of S6: if the number of evolutionary generations exceeds 25 or the error of the lane line recognition error rate of two adjacent tests is less than 0.0001, the search ends; otherwise, it returns to S2.

[0076] After the extreme scene search converges, the complexity of all scenes in the search process and the error detection rate of the lane line algorithm are statistically analyzed. Figure 6 As shown in the figure, it can be seen that in general, the higher the scene complexity, the greater the error detection rate of the lane line algorithm, which shows that the scene complexity index can reflect the extremeness of the scene. The comparison results with typical combination test methods such as Allpairs, AETG and PICT are shown in the figure. Figure 7 As shown, it can be seen that the present invention can significantly increase the proportion of complex scenes, which is more conducive to discovering extreme scenes.

[0077] A person skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing related hardware through a program. The program can be stored in a computer-readable storage medium. When the program is executed, the steps of the method can be implemented. The storage medium, such as ROM / RAM, a disk, an optical disk, etc.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. An extreme scene search method based on an improved genetic algorithm, characterized in that: It includes the following steps: S1. Set the number of scenes parameter N, convergence conditions, and the number of multiple mutation populations N p , a scene parameter value set, and then randomly generate 2N test scenes within the scene parameter value set; S2, treat each scene as a separate population, pair them randomly in pairs, and perform full gene crossover to obtain the crossover offspring population; S3, performing a first screening in the crossover offspring population by using the scene complexity of the population to obtain a crossover optimal offspring population; S4, performing multiple mutation operations on the selected preferred progeny populations to obtain multiple groups of mutant progeny populations; S5. Perform a second screening among multiple groups of mutant offspring populations based on the scene complexity of the population to obtain a mutant optimal offspring population; S6. Use the mutation-selected offspring population to generate test scenarios to test the autonomous driving system, and select the most extreme scenario at the moment according to the elite retention strategy; wherein the elite retention strategy means that the best solution found is retained in each iteration; S7: Determine whether the convergence condition is met according to the test result. If so, the search ends; otherwise, return to S2; In step S2, 2N scenes are randomly paired in pairs, and for each pair of scenes, all N scenes that can be cross-operated are c Single-point crossover mutation is performed at each gene position to obtain N×N c A set of alternative offspring, wherein the single-point crossover mutation refers to exchanging all genes on the right side of the two parental chromosomes for the selected gene crossover point; In step S3, in N×N c Calculate the complexity of the scenes in each group of candidate children, and get the overall complexity of each group of children by summing them up. Then, perform the first screening from high to low according to the overall complexity, and select N groups of scenes. The calculation method of the overall complexity D(X) is: In the formula, the value of each factor x i Construct a test scenario X = {x1,…,x m }, To construct the discrete value points of scene influencing factors using the hierarchical analysis method, for The importance of The value of is calculated through expert scoring method or entropy weight method.

2. The extreme scene search method based on improved genetic algorithm according to claim 1, characterized in that: In step S4, N p The basic bit mutation operation is performed to obtain N p A mutation population; wherein the basic bit mutation operation refers to performing a mutation operation on a certain bit or several bits of genes randomly specified in the individual coding string.

3. The extreme scene search method based on improved genetic algorithm according to claim 2 is characterized by: In step S5, N p The mutant population is screened for the second time, that is, the complexity of each scene in the mutant population is calculated, the overall complexity of the population is obtained by summing up, and then the population with the highest overall complexity is selected as the mutant offspring. The overall complexity calculation method of the second screening is the same as that of the first screening process.

Images