CN117111578A - Automatic driving system detection blind area guiding fuzzy test method and system - Google Patents

Abstract

A method and a system for guiding fuzzy test of a detection blind area of an automatic driving system comprise the steps of randomly generating a plurality of initial scenes, running in a simulator, respectively calculating fitness scores, and carrying out variation on the instantaneous speed and the variation condition of each time node of a traffic participant in the scenes according to the corresponding score condition of each scene; the mutated scenes are used as new use cases to run in a simulator, new fitness scores are obtained, mutation is conducted again based on the new scores, after the mutation is repeated for a plurality of times, scenes with scores larger than the average value are selected for local search, and scenes with scores smaller than the average value are randomly restarted; calculating the similarity among the scenes of the scenes with the scores smaller than the average value, keeping a plurality of scenes with the lowest similarity to update the initial scene, and carrying out the steps again; and (4) changing the parameters in the scene with higher fitness again, and reserving the scene with higher score before and after the change as an output scene. The invention solves the problems of single generated scene, low test efficiency and the like.

Description

A fuzzy testing method and system for automatic driving system detection blind spot guidance

Technical field

The invention belongs to the technical field of intelligent vehicle road systems, and specifically relates to an automatic driving system detection blind spot guidance fuzzy testing method and system.

Background technique

With the rapid development of autonomous driving technology, the safety performance testing technology of autonomous driving systems is also constantly innovating and developing. Early autonomous driving system testing technologies mainly focused on vehicle performance, sensor accuracy, and electronic equipment reliability. As the complexity and criticality of autonomous driving systems continue to increase, testing technologies on functional safety, algorithm robustness, etc. It is also becoming more and more important. External factors such as changing weather conditions, complex traffic environments, and diverse driving tasks all pose new challenges to autonomous vehicle testing. In the early stages of the development of autonomous driving systems, actual vehicle testing is usually used to verify the safety performance of the autonomous driving system. However, actual vehicle testing is mainly suitable for testing and verifying relatively single types of functions. As the autonomous driving level of autonomous vehicles increases, the requirements for test scenarios have become complex and unpredictable, and real vehicle testing can no longer cover all situations it may face.

Simulation testing has become a method favored by various enterprises and research institutions due to its rich testing scenarios, low testing cost, high testing efficiency, and high scenario repeatability. Compared with real vehicle testing, simulation testing can discover functional problems of autonomous vehicles at a lower cost. In particular, it can easily carry out "test-driven development", that is, test results participate in iterative training of algorithms, and has become the most promising method. One of the simulation testing methods.

Fuzz testing is the mainstream and effective software defect detection method in current simulation testing. It is an automatic software testing technology based on defect injection that detects software defects by continuously providing unexpected input to the program under test and monitoring whether the program appears abnormal. It uses a large number of automatically generated test cases as input to the application to discover possible security flaws in the application. Currently, fuzzy testing technology is used in research on safety performance testing of autonomous driving systems. It mainly uses different mutation strategies when generating test scenarios to obtain key violation scenarios that have a greater impact on the safety performance of autonomous driving systems, thereby reducing test cases. quantity to improve testing efficiency.

Contents of the invention

The purpose of the present invention is to provide an automatic driving system detection blind zone guidance fuzzy testing method and system in order to solve the problems of single scene generation and low test efficiency of the key scene search method in the above-mentioned prior art. The simulation simulator is represented by an array. The specific traffic scene is abstracted into a test case, and a fitness function oriented by the sensor detection blind area is designed to represent the safety of each test case. Multiple use cases can be grouped, mutated, iterated, etc., with the ability to search and include Capability of complex test scenarios with multiple vehicles.

In order to achieve the above objects, the present invention has the following technical solutions:

A fuzzy testing method for automatic driving system detection blind spot guidance, including the following steps:

According to the relevant semantic restrictions of the simulator, multiple initial scenes are randomly generated. The scenes include a main vehicle with a built-in autonomous driving system under test and several traffic participants. The duration of each scene and the interval time node are set, and the The instantaneous speed and lane changing conditions of traffic participants at each time node are used as variable parameters;

Run each initial scenario in the simulator, and calculate the fitness score through a pre-designed fitness function describing the safety situation between vehicles in the scenario based on the blind spot occlusion of the main vehicle sensor detection area. According to the corresponding score of each initial scenario, Scenarios with higher scores and poor safety performance are retained, and the instantaneous speed and lane changing conditions of traffic participants at each time node in the scene are mutated;

Run the mutated scenario queue in the simulator as a new use case to obtain a new fitness score, and mutate again based on the new score. After the above process is repeated several times, the mean value of the fitness scores of all scenarios is calculated and selected. Scenes with scores greater than the average are searched locally, and other scenes with scores less than the average are randomly restarted;

After a random restart, calculate the similarity between scenarios for scenarios with fitness scores less than the mean, and retain the scenarios with the lowest similarity as initial scenarios. After updating the initial scenarios, perform the aforementioned steps again;

Based on the parameters in the scenario with higher fitness, mutate again, run the mutated scenario in the simulator, calculate the fitness score, compare the scores of the two scenarios before and after the mutation, and retain the higher one as the output scenario.

As a preferred solution, the variable parameters define a value range in the scene, and the speed of the main vehicle is controlled at a constant speed of 30km/h, and the speed of traffic participants is controlled within the range of 20km/h to 120km/h. The participant's lane changing situation is represented by 0, 1, and 2. 0 represents going straight without changing the lane, 1 represents changing lanes to the left, and 2 represents changing lanes to the right. It also constrains the behavior of all vehicles in the scene and eliminates invalid scenes.

As a preferred solution, the fitness function describing the safety situation between vehicles in the scene is designed as follows based on the blind area occlusion of the main vehicle sensor detection area: if the main vehicle is following the vehicle in the same direction, let the vehicle The length is l and the width is d. Taking the front midpoint of the rectangle of the main vehicle and the lower left corner of the rectangle of the preceding vehicle as the reference point, assuming that the lateral distance to the preceding vehicle is x and the longitudinal distance is y, within the detection range of the detector, the leading vehicle is opposite The blind spot angle S caused by the occlusion of the main vehicle is S. When the vehicle is driving, the closer other vehicles in the scene are to the main vehicle, the larger the blind spot angle S is. As the relative positions between vehicles change, the occlusion angle S is calculated in different ways. .

As a preferred solution, as the relative positions between vehicles change, the occlusion angle S is calculated as follows:

When 0<x≤y-d:

When y-d<x≤y:

When y＜x＜y+l:

As a preferred solution, when two or more non-host vehicles appear in the main vehicle sensor detection area at the same time, the two vehicles simultaneously block the detection of the main vehicle sensor. If the blind spot angles of these vehicles overlap, it is considered that the corresponding The scene is a dangerous scene, otherwise it is safe;

The calculation expression of blind zone coincidence angle S _b is as follows:

S _b =S _area1 ∩S _area2

The blind zone coincidence angle S _b has the following relationship: when S _b >0, it is dangerous, and when S _b =0, it is safe;

The angular range of each vehicle's blind spot is defined as S _area . According to the calculation result of S, the calculation expression of S _area is as follows:

When S _b is larger, the blocking situation detected by the main vehicle sensor by the vehicle in front is more serious, and the distance between the vehicles is also closer; conversely, when the blind spot angle overlap ratio is higher, but the distance between the vehicles is farther, And the proportion of blind spot angle overlap is low, but the distance between vehicles is closer. When these two situations occur, the danger of the corresponding scene is low, and the value of S _b is also low.

As a preferred solution, the mutated scenario queue is run in the simulator as a new use case to obtain a new fitness score, and is mutated again based on the new score. After the above process is repeated several times, the calculation The average of the fitness scores of all scenes, select scenes with scores greater than the average for local search, and randomly restart scenes with scores less than the average. The steps include:

To maximize the discovery of highly adversarial dangerous test scenarios within a given search _budget , define b _t as the set of adversarial scenarios discovered when the input x _t ∈ The expression is as follows:

y _t =max{b ₁ ,b ₂ ,b ₃ ……,b _t }

Using the fuzzy search method based on genetic algorithm, the ultimate goal of fuzzy search is to maximize the various adversarial scenarios discovered, which corresponds to the need to find the optimal individual among the offspring in the genetic algorithm;

In the genetic algorithm, the individual's fitness score is used to determine whether the individual's genes should be retained or eliminated; the fitness function is designed based on the vehicle domain knowledge and scene situation. The fitness function is based on the vehicle-related parameters in the scene. Calculate within time steps the maximum distance α that the main vehicle EV can move without colliding with other vehicles NPC, as well as the occlusion rate β of the main vehicle radar by other vehicles in the scene, and the fitness function at a certain time t in the scene as follows:

f _t (EV,NPC)=α _t +β _t

If a vehicle collides in the scene, an additional constant c will be added as the collision compensation of the fitness function:

f _t (EV,NPC)=α _t +β _t +c

During the scene simulation process, the fitness score is measured once at each time step t. The larger the f _t measured at time step t, the higher the risk of vehicle conflict at time t; according to the dynamic changes of the scene , the fitness score of the scene will increase or decrease after each time step; the calculation expression of the fitness score of the scene is as follows:

Chromosome crossover is designed as an exchange operation of NPCs in the scene. One NPC is randomly selected in each of the two scenes in each generation, and the movement status of the NPC during the running time of the two scenes is exchanged to complete the chromosome crossover operation; Chromosome mutation is An NPC is randomly selected during the genetic process, and its vehicle driving status at a certain moment is randomly changed.

As a preferred solution, the purpose of the scene selection process is to eliminate unsuitable individuals from each generation of scenes and obtain individuals with higher fitness. The roulette selection method is used to achieve this process. If the fitness of the scene The higher it is, the easier it is to choose this scene; in the roulette selection method, the selection probability of an individual is proportional to its fitness. The greater the fitness, the greater the selection probability; Suppose the fitness value of a certain part x _i Expressed as f(xi ₎ , the corresponding probability of being selected is p( _xi ), and the cumulative probability is q( _xi ). The calculation expression is as follows:

As a preferred solution, in the step of calculating the similarity between scenarios for scenarios where the fitness score is less than the mean, and retaining the scenarios with the lowest similarity as the initial scenarios, a random restart mechanism is used to avoid falling into the local minimum. excellent solution;

When the scene fitness score during the iteration process does not improve over time, a random restart is performed at this time;

By comparing the fitness score of the current scene with the average fitness score of previous generations, if the current fitness score is less than the average fitness score of previous generations, random initialization is performed based on the previously recorded NPC trajectories in each scene. In the scene of NPC trajectories, the scene with the lowest similarity to the scene that has been run is selected as the new input of the genetic algorithm after random restart; the similarity is calculated by calculating the Euclidean distance between the NPC trajectories in the candidate scenes and the NPC trajectories in the previous scenes. This is achieved by distance. The larger the distance, the lower the similarity between scenes; the scene similarity is calculated as follows:

S _t ^npc,1 represents the vehicle status of NPC1 at scene t. The vehicle status information includes the position information L _t ^npc,1 and speed information V _t ^npc,1 of each vehicle, and is calculated according to the following expression:

S _t ^npc,1 =L _t ^npc,1 + V _t ^npc,1

use Represents the Euclidean distance of different NPCs at time t. Assuming that the position information of the same NPC in different test cases can be expressed as L _t ^npc,1′ and L _t ^npc,1 , then the scene similarity of Scenario1 and Scenario2 Calculate according to the following expression:

As a preferred solution, it also includes measuring the effectiveness of test cases and the completeness of tests through coverage indicators;

Use two parameters as the parameter space composed of horizontal and vertical coordinates to describe the scene coverage. One parameter is the Euclidean distance between the vehicle under test and the vehicle with occlusion, and the other parameter is the proposed occlusion parameter;

Assume that a total of n test cases have been executed, and the parameters of the i-th test case T _i are (A _i , B _i ) covering a small part of the parameter space. Then the cumulative coverage Cov of the parameter space is expressed as all test cases T The ratio of the covered parameter space area to the total parameter space area Area, the expression is as follows:

Among them, the coverage immunity of each test case is obtained by calculating the area of the rectangle. Assume that the parameter area covered by the i-th test case _Ti is the rectangle (a _i , b _i ) to (A _i , B _i ), then its The coverage area is expressed as:

Area(T _i )=(A _i -a _i )×(B _i -b _i )

Then the cumulative coverage Cov is:

An automatic driving system detection blind spot guidance fuzzy testing system, including:

The initial scene generation module is used to randomly generate multiple initial scenes based on the relevant semantic restrictions of the simulator. The scene contains a main vehicle with a built-in autonomous driving system under test and several traffic participants. The duration of each scene is set. and interval time nodes, and take the instantaneous speed and lane changing conditions of traffic participants at each time node as variable parameters;

The scene queue mutation module is used to run each initial scenario in the simulator and calculate the fitness score through a pre-designed fitness function describing the safety situation between vehicles in the scene based on the blind spot occlusion of the main vehicle sensor detection area. According to Each initial scene corresponds to the score, and scenes with higher scores and poor safety performance are retained, and the instantaneous speed and lane changing conditions of traffic participants at each time node in the scene are mutated;

The local search module is used to run the mutated scene queue as a new use case in the simulator, obtain a new fitness score, and mutate again based on the new score. After the above process is repeated several times, all scene adaptations are calculated. The mean value of the degree score, select the scene with a score greater than the mean value for local search, and randomly restart the other scenes with a score less than the mean value;

The initial scene update module is used after a random restart to calculate the similarity between scenarios for scenarios with fitness scores less than the mean, retain the lowest similarity scenarios as initial scenarios, and perform the aforementioned steps again after updating the initial scenarios;

The output scene selection module is used to mutate again based on the parameters in the scene with higher fitness, run the mutated scene in the simulator, calculate the fitness score, compare the scores of the two scenes before and after the mutation, and retain the higher of one as the output scene.

Compared with the prior art, the present invention at least has the following beneficial effects:

According to the relevant semantic restrictions of the simulator, multiple initial scenes are randomly generated. The scenes include a main vehicle with a built-in autonomous driving system under test and several traffic participants. The specific traffic scenes represented by arrays in the simulation simulator are abstracted. For a test case, each initial scenario is run in the simulator, and a fitness function oriented by the sensor detection blind area is designed to represent the safety of each test case. The present invention combines the genetic algorithm and designs a method based on Detect the fitness function of the blind area to control the generation of test cases, and discover highly adversarial dangerous test scenarios to the maximum extent within a given search budget. A fuzzy search method based on genetic algorithms is used. Starting from some initial inputs, new input, discover new conflicts, and update the input. At the same time, the present invention also uses a random restart mechanism to avoid falling into the local optimal solution. The autonomous driving system makes decisions and controls based on sensors obtaining environmental information. If obstructions hinder the sensor's ability to obtain information, the autonomous driving system may not be able to make correct decisions and controls, leading to accidents. The testing method of the present invention can effectively search for typical test cases that cause the smart car to lose control or collide with other vehicles.

Furthermore, in order to improve testing efficiency, the present invention proposes the concept of scene vehicle behavior constraints to reduce the generation of invalid test cases during the fuzz testing process, and combines it with the fitness function to effectively improve testing efficiency.

Description of drawings

In order to more clearly illustrate the technical solutions in the application of the present invention, the drawings needed to be used in the description of the application will be briefly introduced below. The drawings in the description below are only some embodiments of the application. For those of ordinary skill in the art, As far as workers are concerned, other drawings can also be obtained based on these drawings without exerting creative work.

Figure 1 is an overall flow chart of the blind spot guidance fuzzy testing method for automatic driving system detection according to the embodiment of the present invention;

Figure 2 is a schematic diagram of a vehicle passive collision scenario according to the embodiment of the present invention;

Figure 3 is a schematic diagram of a non-host vehicle conflict scenario according to an embodiment of the present invention;

Figure 4 is a schematic diagram of the front vehicle blocking the sensor detection area of the main vehicle according to the embodiment of the present invention;

Figure 5(a) is a schematic diagram of the front vehicle blocking the sensor detection area of the main vehicle when it is in position 1 according to the embodiment of the present invention;

Figure 5(b) is a schematic diagram of the front vehicle blocking the sensor detection area of the main vehicle when it is in position 2 according to the embodiment of the present invention;

Figure 5(c) is a schematic diagram of the front vehicle blocking the sensor detection area of the main vehicle when it is in position 3 according to the embodiment of the present invention;

Figure 6(a) is a schematic diagram of the overlap angle of the blind spots when two non-host vehicles are in position 1 according to the embodiment of the present invention;

Figure 6(b) is a schematic diagram of the overlap angle of the blind spots when two non-host vehicles are in position 2 according to the embodiment of the present invention;

Figure 6(c) is a schematic diagram of the overlap angle of the blind spots when two non-host vehicles are in position 3 according to the embodiment of the present invention;

Figure 7 is a schematic diagram of chromosome crossover according to an embodiment of the present invention;

Figure 8 is a schematic diagram of chromosome variation according to an embodiment of the present invention;

Figure 9(a) is a schematic diagram of scene 1 under the same occlusion angle according to the embodiment of the present invention;

Figure 9(b) is a schematic diagram of scene 2 at the same occlusion angle according to the embodiment of the present invention;

Figure 10 is a schematic diagram of the final test scenario of the embodiment of the present invention;

Figure 11 is a schematic diagram of the relationship between occlusion and collision according to the embodiment of the present invention;

Figure 12 is a diagram showing the proportion of detection blind areas at different stages between the method according to the embodiment of the present invention and other methods;

Figure 13(a) Comparison of the number of dangerous scene searches when the fuzzer searches the same algebra;

Figure 13(b) Comparison of the average time of dangerous scene search when the fuzzer searches the same algebra.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be described and illustrated below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without any creative work shall fall within the scope of protection of this application.

Please refer to Figure 1. The blind spot guidance fuzzy testing method for automatic driving system detection according to the embodiment of the present invention includes the following steps:

Step 1: According to the relevant semantic restrictions of the simulator, 4 initial scenes are randomly generated. The scenes include a main vehicle with a built-in autonomous driving system under test and 2 traffic participants. The initial positions of all vehicles are fixed at a lane width of 3.7 meters of three-lane highway. Assume that the duration of each scene is 9 seconds, take 3 seconds as a time node, and take the instantaneous speed and lane change of traffic participants at each time node as variable parameters, as shown on the left side of Figure 1: Visualization of each parameter in the scene and its meaning and scene.

Step 2: After obtaining 4 initial scenes, a fitness function describing the safety situation between vehicles in the scene is designed based on the blind spot occlusion of the main vehicle sensor detection area, and the 4 scenes are run in the simulator, and their adaptations are calculated respectively. degree score, and then according to the corresponding scores of each scene, scenarios with higher scores and poor safety performance are retained, and the instantaneous speed and lane changing conditions of traffic participants at each time node in the scene are mutated.

Step 3. The mutated scenario queue is run in the simulator as a new use case to obtain a new fitness score. It is mutated again based on the new score. After the above process is repeated 5 times, the mean value of the fitness scores of all scenarios is calculated. , select scenes with scores greater than the mean for local search, and other scenes with scores less than the mean for random restart.

Step 4: Restart randomly. For scenarios with fitness scores less than the mean, calculate the similarity between the scenarios, and retain the 4 scenarios with the lowest similarity as the initial scenarios to repeat all the processes in steps 2 to 4.

Step 5. Local search, that is, based on the parameters in the scene with higher fitness, mutate it again, run the mutated scene in the simulator, calculate its fitness score, and compare the scores of the two scenes before and after the mutation. Keep the higher one as the output scene.

In a possible implementation, when generating test cases in step 1, in order to ensure that each test case is valid in the scenario, the value range of the variable parameters of the test scenario is defined to control the speed of the main vehicle at a constant speed. 30km/h, the speed of traffic participants is controlled within 20km/h-120km/h. The lane changing status of traffic participants is represented by 0, 1, 2. 0 means going straight without changing the lane, 1 means changing lanes to the left, and 2 Means changing lanes to the right. In addition, the behavior of all vehicles in the scene is constrained, thereby eliminating a large number of invalid scenes in the fuzz test and improving testing efficiency.

The main purpose of this invention is to generate collision scenarios between vehicles. However, defining effective collision scenarios is a difficult problem. Simply limiting the parameters of the scenario will lead to a variety of different situations, but the goal of scene search should be to discover Scenarios that are consistent with real situations and meaningful for safety performance testing of autonomous driving systems. According to the scenario test requirements and relevant traffic regulations, the constraints described in Figures 2 and 3 are defined.

As shown in Figure 2, the main vehicle suddenly encountered a collision with a vehicle from the side and rear during normal driving, resulting in a conflict scene. According to relevant traffic regulations, the rear vehicle bears full responsibility for this conflict. In the embodiment of the present invention, the collision behavior occurring on the side and rear of the host vehicle is defined as a vehicle behavior constraint. If a scene conflict is caused by such behavior, it will be regarded as an invalid scene.

As shown in Figure 3, in this scenario, two non-host vehicles from the front of the host vehicle collided, but it did not have any impact on the driving of the host vehicle, and because in autonomous driving tests, there is usually no relevant setting for the non-host vehicle. control algorithm, so most collisions between non-host vehicles do not conform to real traffic scenarios.

In one possible implementation, the fitness function described in step 2 takes the car-following scene traveling in the same direction as an example to calculate in real time the main vehicle's detection of the preceding vehicle within the sensor range. As shown in the car following scene shown in Figure 4, the main car is following the car in the same direction. The sensor device equipped by the main car is a lidar, where the radar parameters are: θ = 90°, R = 100m.

There are differences in the occlusion of the host vehicle by different vehicle positions, so the calculation of the host vehicle's detection blind spots in the scene is also different. Suppose the length of the vehicle is l and the width is d. At the same time, for the convenience of calculation, each vehicle is regarded as a cuboid with the same length and width. Taking the front midpoint of the main vehicle rectangle and the lower left corner of the front vehicle rectangle as the reference points, let The lateral distance to the front vehicle is x, and the longitudinal distance is y. The blind spot angle caused by the front vehicle blocking the main vehicle within the detection range of the detector is defined as S, as shown in Figure 4.

When the vehicle is driving, the closer other vehicles in the scene are to the main vehicle, the larger the blind spot angle S will be. As the relative positions between vehicles change, the occlusion angle S is calculated differently. As shown in Figure 5(a) to Figure 5(c), when the front vehicle is at different positions in the detection range, the calculation method of the occlusion angle S is also different. The specific calculation is as follows: formulas (1), (2), (3) ) shown.

When 0<x≤y-d:

When y-d<x≤y:

When y＜x＜y+l:

When two or more non-main vehicles appear within the radar detection range of the main vehicle at the same time, and the main vehicle detector is blocked by the two vehicles at the same time, if the blind spot angles of these vehicles overlap, the scene is considered to be a dangerous scene, and vice versa. Safety. The calculation formula of blind zone coincidence angle S _b is as follows:

S _b ＝S _area1 ∩S _area2 (4)

The blind zone coincidence angle S _b has the following relationship. When S _b >0, it is dangerous, and when S _b =0, it is safe. The angular range of each vehicle's blind spot is defined as S _area . According to the calculation result of S, the calculation of S _area is as shown in formula (5):

As shown in Figure 6(a) to Figure 6(c), when S _b is larger, it means that the situation of the front vehicle to the main vehicle sensor is more serious at this time, and at the same time, the distance between the vehicles is also closer. On the contrary, when the proportion of overlapping blind spot angles is high, but the vehicle distance is far away, and when the proportion of overlapping blind spot angles is low, but the vehicle distance is close, the danger of the scene is low, and S _b The value is also lower.

In a possible implementation, in step three, the fundamental purpose is to discover highly adversarial dangerous test scenarios to the maximum extent possible within a given search budget. This can be understood as an optimization problem, how to maximize the number of adversarial scenes Y within a fixed number of iterations T from the space X of all valid input scenes. Therefore, the embodiment of the present invention defines b _t as the set of adversarial scenarios discovered when the number of fuzzy iterations of input x _t ∈X is t times. Then the adversarial scenario y _t can be expressed as follows:

y _t =max{b ₁ ,b ₂ ,b ₃ ……,b _t } (6)

Since the range of the input space X is too large, the exhaustive method cannot be used for search optimization. Furthermore, search focus needs to be identified and placed on promising areas to optimize search efficiency in adversarial scenarios. The embodiment of the present invention uses a fuzzy search method based on genetic algorithm. Starting from some initial inputs, fuzzy searchers tend to select new inputs and discover new conflicts to generate further inputs. The fuzzy search algorithm proposed in the embodiment of the present invention includes the following three parts: genetic algorithm, random restart module and local search module. The pseudocode of the algorithm is as follows:

The ultimate goal of fuzzy search is to maximize the discovery of various adversarial scenarios, which corresponds to the need to find the optimal individual among the offspring in the genetic algorithm. However, due to less input of collision scenarios, more specific violation guidance is needed to help the host vehicle cause corresponding conflicts. For example, in order to obtain a scenario where a vehicle collides with a pedestrian, a fitness function is needed to help lead to a scenario of a traffic violation between a vehicle and a pedestrian.

In a genetic algorithm, an individual's fitness score is used to determine whether the individual's genes should be retained or eliminated. For testing purposes, the embodiment of the present invention chooses to design a fitness function based on vehicle domain knowledge and scene situations. The fitness function of the embodiment of the present invention calculates the maximum distance that the main vehicle (EV) can move without colliding with other vehicles (NPC) in each time step based on the relative position of the vehicle in the scene and vehicle speed and other relevant parameters. α, and the occlusion rate β of other vehicles in the scene to the main vehicle radar. Then the calculation formula of the fitness function at a certain moment t in the scene is:

f _t (EV,NPC)＝α _t +β _t (7)

If a vehicle collides in the scene, an additional constant c will be added as the collision compensation of the fitness function, with the purpose of retaining more dangerous scenes during the scene evolution process.

f _t (EV,NPC)＝α _t +β _t +c (8)

During the scene simulation process, the embodiment of the present invention measures the fitness score every 0.1 s. The larger the f _t measured at time step t, the higher the risk of vehicle conflict at time t. According to the dynamic changes of the scene, the fitness score of the scene will increase or decrease after each time step. The fitness score of the scene is as follows:

The genetic mutation process in nature is relatively complex. After many simplifications, the genetic algorithm finally simulated only two processes: crossover and mutation.

In the embodiment of the present invention, chromosome crossover is designed as an exchange operation of NPCs in the scene. One NPC is randomly selected in each of the two scenes in each generation. By exchanging the movement status (vehicle speed and driving direction) of the NPC during the running time of the two scenes, This completes the crossover operation of chromosomes. As shown in Figure 7.

Chromosome mutation is to randomly select an NPC during the genetic process and randomly change its vehicle driving status at a certain moment. The specific operation is shown in Figure 8.

In one possible implementation, in step three, the scenario selection process is to eliminate unsuitable individuals from each generation of scenarios to obtain individuals with higher fitness. The present invention uses a roulette wheel selection method to implement this process. If the adaptability of the scene is higher, the scene will be selected more frequently. In this method, the selection probability of each individual is proportional to its fitness. The greater the fitness, the greater the selection probability. However, in actual roulette selection, individual choices are often not based on individual selection probabilities, but on cumulative probabilities.

Obviously, in the process of individual selection, the area with the highest proportion of the population has the greatest probability of being selected because it accounts for the highest proportion of the population. But at the same time, some scenarios with low fitness scores can also survive the selection process. The advantage of this is that it enriches the gene pool of the population and avoids falling into local optimality during the evolution process. The probability of each part being selected is proportional to its fitness value. Suppose the fitness value of a certain part x _i is expressed as f(xi ₎ , the probability of this part being selected is p( _xi ), and the cumulative probability is q(xi ₎ . The corresponding calculation formula is as follows:

In a possible implementation, in step 4, the embodiment of the present invention also uses a random restart mechanism to avoid falling into a local optimal solution. When the scene fitness score during the iteration process does not improve over time, a random restart module will be executed. By comparing the fitness score of the current scene with the average fitness score of the previous five generations, if it is found that the current fitness score is less than the average of the previous five generations, then based on the previously recorded NPC trajectory in each scene, at 1000 Among the scenes with randomly initialized NPC trajectories, the scene with the lowest similarity to the previously run scenes is selected as the new input of the genetic algorithm after random restart. The similarity is achieved by calculating the Euclidean distance between the NPC trajectory in the candidate scene and the NPC trajectory in the previous scene. The larger the distance, the lower the similarity between the scenes. The relevant definitions of scene similarity are as follows:

S _t ^npc,1 represents the vehicle status of NPC1 at scene t. The vehicle status information includes the position information L _t ^npc,1 and speed information V _t ^npc,1 of each vehicle. Equation 11 is as follows:

S _t ^npc,1 =L _t ^npc,1 +V _t ^npc,1 (11)

use Represents the Euclidean distance of different NPCs at time t. Assuming that the position information of the same NPC in different test cases can be expressed as L _t ^npc,1′ and L _t ^npc,1 , then the scene similarity of Scenario1 and Scenario2 As shown in the following formula:

On this basis, the present invention proposes a new coverage index to measure the effectiveness of test cases and the completeness of tests.

The autonomous driving system obtains environmental information based on sensors for decision-making and control. If obstructions hinder the sensor's ability to obtain information, the autonomous driving system may not be able to make correct decisions and control, leading to an accident.

Therefore, it is very important to test the performance of the autonomous driving system in occlusion situations, because this situation may cause the vehicle to be unable to recognize, track or avoid other vehicles, pedestrians, obstacles, etc., thus endangering the lives of drivers and passengers.

As shown in Figure 9(a) and Figure 9(b), the occlusion angle θ ₁ = θ ₂ in the two scenes, but it can be clearly found that the occlusion situations in the two scenes are completely different. In scene 1, the main vehicle is blocked due to the distance Closer, partially obscured vehicles can be detected. In scene 2, because the main vehicle is far away, most areas of the frontmost vehicle cannot be detected by the main vehicle.

Through further analysis of the above two scenes, it can be found that the dangers of the two scenes will be significantly different when faced with different behaviors of vehicles in the scene. For example, when the car 2 in front changes lanes to the left, scene 1 Although there are fewer blocked angles, because the distance between the vehicles is closer, there is less room for the host vehicle to react, making danger more likely to occur.

In contrast, when the main vehicle in scenario 2 faces the same situation, it has more time to take measures because it is far away from the leading vehicle 2. On the contrary, when the front car 1 in the scene drives away to the right, in scene 1 because the main car and the front car 2 are not in the same lane, there is no risk of collision. In scene 2, although the main car and the front car 2 are not in the same lane, there is no risk of collision. The vehicle 2 in front is far away, but still in the same lane, so when the vehicle 1 in front drives away, the main vehicle is at risk of colliding with the vehicle 2 in front.

In view of this, the present invention uses two parameters as a parameter space composed of horizontal and vertical coordinates to describe the scene coverage. One parameter is the Euclidean distance between the vehicle under test and the vehicle with occlusion, and the other parameter is the Euclidean distance proposed in the embodiment of the present invention. occlusion parameters. Assume that a total of n test cases are executed in the experiment, and the parameters of the i-th test case T _i are (A _i , B _i ), covering a small part of the parameter space. Then the cumulative coverage Cov of the parameter space can be expressed as the ratio of the parameter space area covered by all test cases T to the total parameter space area Area, that is:

Among them, the coverage immunity of each test case can be obtained by calculating the area of the rectangle in which it is located. Assume that the parameter area covered by the i-th test case T _i is a rectangle (a _i , b _i ) to (A _i , B _i ), then its coverage area can be expressed as:

Area(T _i )=(A _i -a _i )×(B _i -b _i ) (15)

Substituting the above formula into the cumulative coverage formula, we get:

The technical effects of the present invention will be further verified and explained through simulation experiments below.

lab environment:

The machine used in the experiment of this invention is: Dell Precison 3640Tower desktop computer; the processor is: Intel i9-10900K; 32G memory; the operating system is 64-bit window10.

In order to ensure the fairness of the experimental results, the same initial scene was selected for different scene generation methods. The simulation software used is the automatic driving toolbox in Matlab R2021b. The toolbox provides driving scenarios, sensors, vehicle functional modules and 3D simulation modules. Based on these modules, various required automatic driving simulation models can be constructed for Simulation of autonomous driving. Use the helperGridBasedPlanningScenario module and Motion Planning in Urban EnvironmentsUsing Dynamic Occupancy Grid Map algorithm provided by the autonomous driving toolbox as the test platform and autonomous driving control algorithm in the virtual safety test to simulate the behavioral logic of the autonomous vehicle in the virtual driving scene. By using The simulated sensor data provided in the helperGridBasedPlanningScenario module is used to perform driving tasks including obstacles, path planning and controlling the vehicle in a given driving scenario.

In order to verify the effectiveness of the method of the present invention, in the experiment, each test needs to iterate the scene 100 times. Finally, after many experiments, a total of 3 typical tests that can cause the smart car to lose control or collide with other vehicles were searched. Example.

Figure 10 is a highly confrontational test scene searched during the experiment. In this scene, the main vehicle is driving in a straight line at a constant speed on the road. When the main car tried to overtake two non-main vehicles in the adjacent lane in the same direction, the yellow car in the right lane changed lanes. The main car was faster and the radar detector was blocked by non-main vehicle 1. , causing the main vehicle to be unable to detect the lane-changing vehicle ahead in time, eventually leading to a collision. Normally, when a human driver encounters a visual blind spot, he or she will reduce the speed appropriately based on the actual situation. Especially when the vehicle in front is found to be intending to change lanes, the human driver will drive more cautiously to avoid accidents. For the automatic driving system, if it is determined that the detection blind spot is large and the speed of the main vehicle is higher than that of surrounding vehicles, the automatic driving system should adopt a conservative driving strategy and reduce the vehicle speed.

In order to verify whether the vehicle collision is related to the blocking of the main vehicle by the vehicle in front, a supplementary experiment as shown in Figure 11 was conducted: by changing the position of the red car so that the red car could not block the main vehicle, it was finally found that the main vehicle could successfully Identify the yellow car and successfully perform avoidance maneuvers.

To verify the effectiveness of the method, it is necessary to verify whether the method of the embodiment of the present invention improves the proportion of the main vehicle detection blind area in the generated input. Use the method of the embodiment of the present invention, AV-fuzzer and random fuzzer to collect each newly generated input during the fuzz testing process. After the fuzz testing is completed, test the proportion of the main vehicle detection blind area in these inputs and compare them. The results are as shown in the figure 12 shown. In Figure 12, the horizontal axis represents the number of iterations, and the vertical axis represents the proportion of the main vehicle detection blind area corresponding to each generation of input. It can be seen from Figure 12 that at the beginning of the test, the proportion of the main vehicle detection blind area generated by the method of the embodiment of the present invention is slightly lower than that of the AV-fuzzer. This is because the method of the embodiment of the present invention generates input randomly in the early stage, and the test method also In the exploration and utilization stage. In the middle and later stages of testing, the blind area ratio of the method of the embodiment of the present invention is significantly higher than that of the AV-fuzzer and the random fuzzer. This is because after the exploration and utilization in the previous stages, the testing tool based on the method of the present invention utilizes the unique characteristics defined by the embodiment of the present invention. The knowledge learned by the fitness function selects modification actions, which increases the possibility that new inputs can produce more occlusions by non-host vehicles on the host vehicle, thereby increasing the proportion of detection blind spots.

In the embodiment of the present invention, AV-fuzzer and random fuzzer are selected for comparative experiments. The target of the test is the built-in path planning algorithm of Matlab2021b. In order to evaluate the overall effect of the test method, the initial scenario and mutation probability of each method are set to the same value. After the same 100 iterative searches, the effectiveness of the three fuzzers is Performance and efficiency are evaluated and represented graphically.

Figure 13(a) and Figure 13(b) show the number of dangerous test cases searched by the three methods when the fuzzer searches the same algebra. The BlindSpotsFuzz proposed in the embodiment of the present invention searched for a total of 9 different dangerous test cases, the AV-fuzzer found 5 dangerous test cases, and the random fuzzer did not find any dangerous test cases. In terms of time, BlindSpotsFuzz takes the least average time to search for a dangerous test case. To sum up, within the same fuzz testing time and number of iterations, BlindSpotsFuzz detected the most dangerous scenarios. Compared with other methods, BlindSpotsFuzz can mutate more dangerous test cases based on collisions, and can eliminate invalid test cases during the fuzz testing process, further improving fuzzy search efficiency. In addition, BlindSpotsFuzzr found that 4 of the 9 dangerous test cases were not detected by other fuzzers, while the 5 test cases found by AV-fuzzer were all detected by BlindSpotsFuzz. This shows that the method of the present invention is an improvement on current smart cars. A powerful addition to fuzz testing.

Another embodiment of the present invention also proposes an automatic driving system detection blind spot guidance fuzzy testing system, including:

The initial scene generation module is used to randomly generate multiple initial scenes based on the relevant semantic restrictions of the simulator. The scene contains a main vehicle with a built-in autonomous driving system under test and several traffic participants. The duration of each scene is set. and interval time nodes, and take the instantaneous speed and lane changing conditions of traffic participants at each time node as variable parameters;

The scene queue mutation module is used to run each initial scenario in the simulator and calculate the fitness score through a pre-designed fitness function describing the safety situation between vehicles in the scene based on the blind spot occlusion of the main vehicle sensor detection area. According to Each initial scene corresponds to the score, and scenes with higher scores and poor safety performance are retained, and the instantaneous speed and lane changing conditions of traffic participants at each time node in the scene are mutated;

The local search module is used to run the mutated scene queue as a new use case in the simulator, obtain a new fitness score, and mutate again based on the new score. After the above process is repeated several times, all scene adaptations are calculated. The mean value of the degree score, select the scene with a score greater than the mean value for local search, and randomly restart the other scenes with a score less than the mean value;

The initial scene update module is used after a random restart to calculate the similarity between scenarios for scenarios with fitness scores less than the mean, retain the lowest similarity scenarios as initial scenarios, and perform the aforementioned steps again after updating the initial scenarios;

The output scene selection module is used to mutate again based on the parameters in the scene with higher fitness, run the mutated scene in the simulator, calculate the fitness score, compare the scores of the two scenes before and after the mutation, and retain the higher of one as the output scene.

Another embodiment of the present invention also provides an electronic device, including: a memory that stores at least one instruction; and a processor that executes the instructions stored in the memory to implement the automatic driving system detection blind spot guidance fuzz testing method of the present invention.

Another embodiment of the present invention also provides a computer-readable storage medium. The computer-readable storage medium stores at least one instruction. The at least one instruction is executed by a processor in an electronic device to implement the automatic operation of the present invention. Driving system detection blind spot guidance fuzzy testing method.

Exemplarily, the instructions stored in the memory may be divided into one or more modules/units, and the one or more modules/units are stored in a computer-readable storage medium and executed by the processor, To complete the automatic driving system detection blind spot guidance fuzzy testing method of the present invention. The one or more modules/units may be a series of computer-readable instruction segments capable of completing specific functions. The instruction segments are used to describe the execution process of the computer program in the server.

The electronic device may be a computing device such as a smartphone, a notebook, a PDA, a cloud server, etc. The electronic device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the electronic device may also include more or less components, or a combination of certain components, or different components. For example, the electronic device may also include input and output devices, network access devices, bus etc.

The processor may be a Central Processing Unit (CPU), or other general-purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), or an off-the-shelf processor. Programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

The memory may be an internal storage unit of the server, such as a hard disk or memory of the server. The memory may also be an external storage device of the server, such as a plug-in hard drive, a smart media card (SMC), a secure digital (SD) card, a flash memory card ( Flash Card), etc. Further, the memory may also include both an internal storage unit of the server and an external storage device. The memory is used to store the computer readable instructions and other programs and data required by the server. The memory may also be used to temporarily store data that has been output or is to be output.

It should be noted that the information interaction, execution process, etc. between the above-mentioned module units are based on the same concept as the method embodiments, and their specific functions and technical effects can be found in the method embodiments section, which will not be discussed here. Repeat.

Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above functional units and modules is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs. Module completion means dividing the internal structure of the device into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be hardware-based. It can also be implemented in the form of software functional units. In addition, the specific names of each functional unit and module are only for the convenience of distinguishing each other and are not used to limit the scope of protection of the present application. For the specific working processes of the units and modules in the above system, please refer to the corresponding processes in the foregoing method embodiments, and will not be described again here.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium. Based on this understanding, this application can implement all or part of the processes in the methods of the above embodiments by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. The computer program When executed by a processor, the steps of each of the above method embodiments may be implemented. Wherein, the computer program includes computer program code, which may be in the form of source code, object code, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device capable of carrying computer program code to the camera device/terminal device, recording media, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), electrical carrier signals, telecommunications signals, and software distribution media. For example, U disk, mobile hard disk, magnetic disk or CD, etc.

In the above embodiments, each embodiment is described with its own emphasis. For parts that are not detailed or documented in a certain embodiment, please refer to the relevant descriptions of other embodiments.

The above-described embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still implement the above-mentioned implementations. The technical solutions described in the examples are modified, or some of the technical features are equivalently replaced; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions in the embodiments of this application, and should be included in within the protection scope of this application.

Claims (10)

1. A fuzzy testing method for automatic driving system detection blind spot guidance, which is characterized by including the following steps: According to the relevant semantic restrictions of the simulator, multiple initial scenes are randomly generated. The scenes include a main vehicle with a built-in autonomous driving system under test and several traffic participants. The duration of each scene and the interval time node are set, and the The instantaneous speed and lane changing conditions of traffic participants at each time node are used as variable parameters; Run each initial scenario in the simulator, and calculate the fitness score through a pre-designed fitness function describing the safety situation between vehicles in the scenario based on the blind spot occlusion of the main vehicle sensor detection area. According to the corresponding score of each initial scenario, Scenarios with higher scores and poor safety performance are retained, and the instantaneous speed and lane changing conditions of traffic participants at each time node in the scene are mutated; Run the mutated scenario queue in the simulator as a new use case to obtain a new fitness score, and mutate again based on the new score. After the above process is repeated several times, the mean value of the fitness scores of all scenarios is calculated and selected. Scenes with scores greater than the average are searched locally, and other scenes with scores less than the average are randomly restarted; After a random restart, calculate the similarity between scenarios for scenarios with fitness scores less than the mean, and retain the scenarios with the lowest similarity as initial scenarios. After updating the initial scenarios, perform the aforementioned steps again; Based on the parameters in the scenario with higher fitness, mutate again, run the mutated scenario in the simulator, calculate the fitness score, compare the scores of the two scenarios before and after the mutation, and retain the higher one as the output scenario. 2. The automatic driving system detection blind spot guidance fuzzy testing method according to claim 1, characterized in that the variable parameter defines a value range in the scene, and the speed of the main vehicle is controlled at a constant speed of 30km/h, and traffic participation Within the speed control range of 20km/h to 120km/h, traffic participants’ lane changing conditions are represented by 0, 1, and 2. 0 represents going straight without changing lanes, 1 represents changing lanes to the left, and 2 represents changing lanes to the right. ; It also restricts the behavior of all vehicles in the scene and eliminates invalid scenes. 3. The automatic driving system detection blind spot guidance fuzzy testing method according to claim 1, characterized in that the fitness function describing the safety situation between vehicles in the scene is as follows according to the blind spot occlusion of the main vehicle sensor detection area. Method design: If the main vehicle follows the vehicle in the same direction, let the length of the vehicle be l and the width be d. Taking the front midpoint of the rectangle of the main vehicle and the lower left corner of the rectangle of the preceding vehicle as the reference point, let the lateral distance to the preceding vehicle be x, the longitudinal distance is y, and the blind spot angle S caused by the vehicle in front blocking the main vehicle within the detection range of the detector is S. When the vehicle is driving, the closer the other vehicles in the scene are to the main vehicle, the larger the blind spot angle S will be. Depending on the change in the relative position between vehicles, the occlusion angle S is calculated in different ways. 4. The automatic driving system detection blind spot guidance fuzzy testing method according to claim 3, characterized in that as the relative positions between vehicles change, the calculation methods of the occlusion angle S are as follows: When 0<x≤y-d: When y-d<x≤y: When y＜x＜y+l: 5. The automatic driving system detection blind spot guidance fuzzy testing method according to claim 3, characterized in that when more than two non-main vehicles appear in the main vehicle sensor detection area at the same time, the two vehicles detect the main vehicle sensor at the same time. Without occlusion, if the blind spot angles of these vehicles overlap, the corresponding scene is considered a dangerous scene, and vice versa, it is safe; The calculation expression of blind zone coincidence angle S _b is as follows: S,＝S _area] S _area2 The blind zone coincidence angle S _b has the following relationship: when S _b >0, it is dangerous, and when S _b =0, it is safe; The angular range of each vehicle's blind spot is defined as S _area . According to the calculation result of S, the calculation expression of S _area is as follows: When S _b is larger, the blocking situation detected by the main vehicle sensor by the vehicle in front is more serious, and the distance between the vehicles is also closer; conversely, when the blind spot angle overlap ratio is higher, but the distance between the vehicles is farther, And the proportion of blind spot angle overlap is low, but the distance between vehicles is closer. When these two situations occur, the danger of the corresponding scene is low, and the value of S _b is also low. 6. The automatic driving system detection blind spot guidance fuzz testing method according to claim 1, characterized in that the mutated scene queue is run in the simulator as a new use case to obtain a new fitness score, based on The new scores are mutated again. After the above process is repeated several times, the mean value of the fitness scores of all scenarios is calculated. Scenarios with scores greater than the mean are selected for local search. Other scenarios with scores less than the mean are randomly restarted. The steps include: To maximize the discovery of highly adversarial dangerous test scenarios within a given search _budget , define b _t as the set of adversarial scenarios discovered when the input x _t ∈ The expression is as follows: y _t =max{b ₁ ,b ₂ ,b ₃ ……,b _t } Using the fuzzy search method based on genetic algorithm, the ultimate goal of fuzzy search is to maximize the various adversarial scenarios discovered, which corresponds to the need to find the optimal individual among the offspring in the genetic algorithm; In the genetic algorithm, the individual's fitness score is used to determine whether the individual's genes should be retained or eliminated; the fitness function is designed based on the vehicle domain knowledge and scene situation. The fitness function is based on the vehicle-related parameters in the scene. Calculate within time steps the maximum distance α that the main vehicle EV can move without colliding with other vehicles NPC, as well as the occlusion rate β of the main vehicle radar by other vehicles in the scene, and the fitness function at a certain time t in the scene as follows: f _t (EV,NPC)=α _t +β _t If a vehicle collides in the scene, an additional constant c will be added as the collision compensation of the fitness function: f _t (EV,NPC)=α _t +β _t +c During the scene simulation process, the fitness score is measured once at each time step t. The larger the f _t measured at time step t, the higher the risk of vehicle conflict at time t; according to the dynamic changes of the scene , the fitness score of the scene will increase or decrease after each time step; the calculation expression of the fitness score of the scene is as follows: Chromosome crossover is designed as an exchange operation of NPCs in the scene. One NPC is randomly selected in each of the two scenes in each generation, and the movement status of the NPC during the running time of the two scenes is exchanged to complete the chromosome crossover operation; Chromosome mutation is An NPC is randomly selected during the genetic process, and its vehicle driving status at a certain moment is randomly changed. 7. The automatic driving system detection blind spot guidance fuzzy testing method according to claim 6, characterized in that the purpose of the scene selection process is to eliminate unsuitable individuals from each generation of scenes and obtain individuals with higher fitness, using The roulette selection method is used to implement this process. If the fitness of the scene is higher, the easier it is to select this scene; in the roulette selection method, the selection probability of an individual is proportional to its fitness, the greater the fitness. , the greater the probability of selection; assuming that the fitness value of a certain part x _i is expressed as f(xi ₎ , the corresponding probability of being selected is p(xi ₎ , and the cumulative probability is q(xi ₎ . The calculation expression is as follows: 8. The automatic driving system detection blind spot guidance fuzzy testing method according to claim 1, characterized in that, for the scenes with fitness scores less than the mean value, the similarity between each scene is calculated, and the scenes with the lowest similarity are retained. As the initial scenario step, a random restart mechanism is used to avoid falling into the local optimal solution; When the scene fitness score during the iteration process does not improve over time, a random restart is performed at this time; By comparing the fitness score of the current scene with the average fitness score of previous generations, if the current fitness score is less than the average fitness score of previous generations, random initialization is performed based on the previously recorded NPC trajectories in each scene. In the scene of NPC trajectories, the scene with the lowest similarity to the scene that has been run is selected as the new input of the genetic algorithm after random restart; the similarity is calculated by calculating the Euclidean distance between the NPC trajectories in the candidate scenes and the NPC trajectories in the previous scenes. This is achieved by distance. The larger the distance, the lower the similarity between scenes; the scene similarity is calculated as follows: S _t ^npc,1 represents the vehicle status of NPC1 at scene t. The vehicle status information includes the position information L _t ^npc,1 and speed information V _t ^npc,1 of each vehicle. It is calculated according to the following expression: S _t ^{npc, 1} = It ^{npc, 1} + V _t ^{npc, 1} use Represents the Euclidean distance of different NPCs at time t. Assume that the position information of the same NPC in different test cases can be expressed as L _t ^npc,1′ and L _t ^npc,1 , then the scene similarity of Scenario1 and Scenario2/ > Calculate according to the following expression: 9. The automatic driving system detection blind zone guidance fuzz testing method according to claim 1, characterized in that it also includes measuring the effectiveness of the test cases and the completeness of the test through coverage indicators; Use two parameters as the parameter space composed of horizontal and vertical coordinates to describe the scene coverage. One parameter is the Euclidean distance between the vehicle under test and the vehicle with occlusion, and the other parameter is the proposed occlusion parameter; Assume that a total of n test cases have been executed, and the i-th test case T _i has parameters (A _i , B _i ) covering a small part of the parameter space. Then the cumulative coverage Cov of the parameter space is expressed as all test cases T The ratio of the covered parameter space area to the total parameter space area Area, the expression is as follows: Among them, the coverage immunity of each test case is obtained by calculating the area of the rectangle. Assume that the parameter area covered by the i-th test case _Ti is the rectangle (a _i , b _i ) to (A _i , B _i ), then its The coverage area is expressed as: Area(T _i )=(A _i -a _i )×(B _i -b _i ) Then the cumulative coverage Cov is: 10. An automatic driving system detection blind spot guidance fuzzy testing system, characterized by including: The initial scene generation module is used to randomly generate multiple initial scenes based on the relevant semantic restrictions of the simulator. The scene contains a main vehicle with a built-in autonomous driving system under test and several traffic participants. The duration of each scene is set. and interval time nodes, and take the instantaneous speed and lane changing conditions of traffic participants at each time node as variable parameters; The scene queue mutation module is used to run each initial scenario in the simulator and calculate the fitness score through a pre-designed fitness function describing the safety situation between vehicles in the scene based on the blind spot occlusion of the main vehicle sensor detection area. According to Each initial scene corresponds to the score, and scenes with higher scores and poor safety performance are retained, and the instantaneous speed and lane changing conditions of traffic participants at each time node in the scene are mutated; The local search module is used to run the mutated scene queue as a new use case in the simulator, obtain a new fitness score, and mutate again based on the new score. After the above process is repeated several times, all scene adaptations are calculated. The mean value of the degree score, select the scene with a score greater than the mean value for local search, and randomly restart the other scenes with a score less than the mean value; The initial scene update module is used after a random restart to calculate the similarity between scenarios for scenarios with fitness scores less than the mean, retain the lowest similarity scenarios as initial scenarios, and perform the aforementioned steps again after updating the initial scenarios; The output scene selection module is used to mutate again based on the parameters in the scene with higher fitness, run the mutated scene in the simulator, calculate the fitness score, compare the scores of the two scenes before and after the mutation, and retain the higher of one as the output scene.