Virtual acceleration test method and system for autonomous driving vehicle based on optimal distribution model

Through the optimal distribution model and importance sampling method, the testing efficiency of high-risk events in virtual scenarios is improved, the problems of long test cycles and low efficiency in the existing technology are solved, and more realistic and efficient safety assessment of autonomous driving cars is achieved. .

CN118446012BActive Publication Date: 2024-11-26JIANGSU UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410596895.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-14
Publication Date
2024-11-26
Estimated Expiration
2044-05-14

AI Technical Summary

Technical Problem

The prior art is difficult to effectively accelerate the high-risk event testing of autonomous vehicles in virtual scenarios, resulting in a long test cycle and the inability to maximize acceleration efficiency. The prior art is unable to accurately characterize the distribution form of complex data, affecting the authenticity and efficiency of the test. .

Method used

Using virtual acceleration testing method based on the optimal distribution model, the maximum likelihood estimation function is constructed to increase high risk by collecting natural driving data, statistics and fitting probability density histograms, and using mixed segmentation functions and optimal importance sampling methods. The sampling probability of events and security testing is performed in the simulation platform.

Benefits of technology

It significantly reduces the number of tests, increases the probability of high-risk events, ensures the credibility and efficiency of the test, significantly accelerates the testing process, and shortens the test cycle.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118446012B_ABST
    Figure CN118446012B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model, including a natural driving data acquisition module, a key variable distribution fitting module, an optimal importance sampling module, and a virtual scene simulation test module. The driving trajectory and motion state of the vehicle are obtained; the probability distribution of key variables is statistically analyzed, and the distribution characteristics of key variables are described by a mixed piecewise function; the original distribution characteristics of key variables are changed by the importance sampling method, the sampling probability of high-risk events is increased, and the optimal distribution for accelerated testing of high-risk events is obtained; the test scene is reconstructed based on the optimal distribution, and the safety test of the autonomous driving vehicle is performed on a simulation test platform, and the probability of occurrence of high-risk events and the acceleration ratio of the test are calculated. The present invention reduces the number of simulation tests while improving the fitting accuracy of the probability distribution through a mixed piecewise function and the optimal importance sampling method, thereby achieving an efficient evaluation of the safety performance of the autonomous driving vehicle.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to autonomous driving technology, and in particular to a method and system for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model. Background Art

[0002] As the level of intelligence increases, autonomous driving systems become more complex, requiring them to deal with uncertain interference in the real world and make more complex decisions. This complexity makes it more challenging to conduct safety assessments of autonomous driving systems.

[0003] Traditional safety assessments are mainly based on the accumulation of mileage in field road tests. Mileage-based safety assessments not only require a massive amount of mileage accumulation and high testing costs, but also fail to conduct extreme high-risk scenario tests, making it difficult to reveal safety hazards of autonomous vehicles in various scenarios. Currently, simulation testing based on virtual scenarios has become the mainstream idea, which can significantly reduce the cost of autonomous driving safety assessments and improve assessment efficiency through simulation time acceleration, cloud platform parallel computing, and accelerated testing methods.

[0004] The core of virtual simulation testing is to construct test demand scenarios in a targeted manner and improve the efficiency of hazard rate assessment for autonomous vehicles.

[0005] In the prior art, CN107421752A can better characterize the data distribution form and achieve a certain acceleration effect by numerically transforming a simple probability distribution function; CN114818381A increases the probability of occurrence of high-risk scenarios through multiple importance sampling, but it requires analysis of the data distribution form and selection of a suitable sampling function, which will increase the complexity of the test; WO2023193196A1 determines high-risk scenarios by clustering driving behavior data.

[0006] It can be seen that for any form of complex data, the accelerated testing scheme of the prior art, the general distribution function cannot well characterize the distribution form of the data through the general probability distribution function, and the data needs to be analyzed and processed, which affects the authenticity of the data. In addition, the accelerated testing scheme of the prior art cannot guarantee the maximization of acceleration efficiency and cannot solve the problem of time-consuming scene repetition. Under the premise of ensuring the accuracy and unbiasedness of the virtual scene, how to increase the test probability of high-risk events with the least test mileage, and solve the problem of ultra-long test cycles caused by time-consuming scene repetition and difficulty in covering high-risk scenes. Summary of the invention

[0007] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide a method and system for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model.

[0008] Technical solution: A virtual acceleration test method for an autonomous driving vehicle based on an optimal distribution model of the present invention comprises the following steps:

[0009] Step 1: Collect natural driving data, collect and record the vehicle's driving trajectory and motion status, including extracting the front vehicle's speed when the front vehicle cuts in V p , Test vehicle speed V c And the distance between the two cars R ;

[0010] Step 2: Count the distances R The probability distribution of R The probability density histogram is then fitted with a mixed piecewise function to obtain the distribution characteristics of the key variables; the mixed piecewise function includes the exponential distribution function, the normal distribution function and the Pareto distribution function;

[0011] Step 3: Use the optimal importance sampling function to change the original distribution characteristics of key variables, increase the sampling probability of high-risk events, take the minimum number of tests as the optimization goal, construct the maximum likelihood estimation function, and obtain the optimal distribution for accelerated testing of high-risk events;

[0012] Step 4: Reconstruct high-risk test scenarios, conduct safety tests on autonomous vehicles in the simulation test platform, and calculate the probability of high-risk events and the acceleration ratio of the test;

[0013] Step 5: Evaluate test efficiency.

[0014] Furthermore, the step 1 collects vehicle video or image information through a camera, collects vehicle distance information through a lidar, and obtains vehicle location information through a GPS.

[0015] Furthermore, the step 2 is based on the distance between the two vehicles obtained in step 1. R The data is counted as a probability density histogram; then the R The probability density histogram is divided into equally spaced segments, and the fitting distribution function with the highest fitting accuracy in each small segment of the histogram is selected as the best fitting function for the small segment of the histogram.

[0016] Furthermore, the specific method for obtaining the best fitting function is as follows:

[0017] Step 1) Set X is a sequence of values ​​at the center of each probability histogram, Y For the sequence X The sequence of median values ​​and corresponding probability density values, is the probability distribution function in the sequenceX The function value at each element in is:

[0018]

[0019]

[0020]

[0021] in, For the i The value of the center position of the probability histogram; for The corresponding probability density value; for The corresponding probability distribution function value; m is the number of bins contained in the histogram, ;

[0022] Step 2), Calculate Y and The absolute error is used to measure the fitting accuracy of each probability distribution function; the function with the highest fitting accuracy in each histogram interval should satisfy the following formula:

[0023]

[0024] In the formula, For the i The best fitting function of the segment; k are the numbers corresponding to the three probability distribution functions, 1, 2 and 3 correspond to the exponential distribution function, the normal distribution function and the Pareto distribution function respectively; For the k The probability distribution function is i The function value of the segment; for X A sequence of the best fitting function values ​​at each element in , where * indicates optimal.

[0025] Furthermore, when the step 3 uses the importance sampling method to generate a probability distribution function with a higher risk probability to replace the original probability distribution function, it is necessary to first use the maximum likelihood estimation function to represent the relationship between the optimal probability distribution function and the original probability distribution function, which is specifically expressed as follows:

[0026]

[0027] In the formula, It is the function with the highest fitting accuracy for natural driving data; is the optimal probability distribution function generated after importance sampling;

[0028] High risk areas are , is the indicator function of high risk areas, The probability of occurrence of high-risk events is:

[0029]

[0030]

[0031] In the formula, n is the number of tests; j For test samples;

[0032] To make exist Convergence is achieved at a confidence level of , the number of tests corresponding to the optimal distribution function n Should meet:

[0033]

[0034] In the formula, is the importance sampling parameter; is the likelihood ratio function; is the confidence level; is the cumulative distribution function value corresponding to the confidence level; is the convergence threshold.

[0035] Furthermore, the original probability distribution function is transformed by importance sampling to improve the sampling probability of high-risk events while ensuring its unbiasedness.

[0036] The measurement transformations corresponding to the original exponential distribution function, normal distribution function, and Pareto distribution function are:

[0037]

[0038] k are the numbers corresponding to the three probability distribution functions, 1, 2 and 3 correspond to the exponential distribution function, the normal distribution function and the Pareto distribution function respectively; , are the coefficients of the three probability distribution functions respectively; is the corresponding acceleration parameter;

[0039] The unbiased constraint is that the original distribution function and the importance sampling function JS The divergence is less than or equal to 0.1, that is:

[0040] .

[0041] Furthermore, the detailed process of step 4 is as follows:

[0042] Step 4.1: Scenario reconstruction, that is, resampling the optimal importance sampling function to generate a new test scenario ; After importance sampling, the new spacing distribution is obtained through the optimal distribution model;

[0043] Step 4.2: Simulation test, that is, build a virtual test platform for the front vehicle cutting-in scenario and As the initial state of the simulation test; control the behavior of the test vehicle through the vehicle control model and record the driving state of the vehicle;

[0044] Step 4.3: Simulation evaluation, that is, based on the collision rate of each test, evaluate the danger probability of the scenario and the test mileage required for the danger probability to converge, thereby evaluating the test efficiency.

[0045] Furthermore, the test efficiency is calculated in The number of tests and test mileage at which the probability of occurrence of high-risk scenarios reaches convergence at a confidence level;

[0046] The test mileage of the autonomous vehicle in natural driving scenarios and acceleration scenarios is compared using the acceleration ratio to evaluate the test efficiency of the hybrid segmented acceleration test method.

[0047] The calculation formula is as follows:

[0048]

[0049]

[0050]

[0051] In the formula, is the speedup ratio of the two test methods; Test mileage for autonomous vehicles in natural driving scenarios; Test mileage for autonomous vehicles in segmented acceleration scenarios; is the frequency of the front vehicle cutting in during natural driving scenarios; The number of tests of autonomous vehicles in natural driving scenarios; The number of tests for autonomous vehicles in mixed segment acceleration scenarios; The moment when the scene of the front car cuts in occurs; For the n The time the vehicle was driven during the test.

[0052] The present invention also discloses a system for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model, comprising a natural driving data acquisition module, a key variable distribution fitting module, an optimal importance sampling module, and a virtual scene simulation testing module;

[0053] The natural driving data acquisition module provides data acquisition for functional scenarios, logical scenarios and specific scenarios for specific testing requirements of the autonomous driving system, collects vehicle video or image information through a camera, collects vehicle distance information through a lidar, and obtains vehicle location information through GPS;

[0054] The key variable distribution fitting module performs probability density statistics on key variables that can assess the danger of a scene, and selects a suitable distribution function to fit its probability density histogram;

[0055] The optimal importance sampling module evaluates the test efficiency of the original three distribution functions, and selects a distribution function with a higher probability of danger and the least number of tests to replace the original distribution function;

[0056] The virtual scene simulation test module resamples the optimal importance sampling function in the optimal importance sampling module to generate a new test scene, and evaluates the safety performance of the test vehicle in the simulation platform.

[0057] Furthermore, the key variable distribution fitting module counts the front and rear distance R of the two vehicles as a probability density histogram and divides it into equally spaced segments, and selects the fitting distribution function with the highest fitting accuracy in each small histogram segment as the best fitting function of the small histogram interval. The best fitting function is selected from the exponential distribution function, the normal distribution function, and the Pareto distribution function.

[0058] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0059] 1. The virtual acceleration test method for an autonomous driving vehicle proposed in the present invention improves the fitting accuracy of the probability density histogram through a mixed segmentation method, more realistically reflects the driving characteristics of natural driving data, and ensures the credibility of the simulation test.

[0060] 2. The virtual acceleration test method for autonomous driving vehicles proposed in the present invention improves the probability of high-risk events in natural driving data while ensuring data unbiasedness by designing an importance sampling form, significantly reduces the number of tests, and is of great significance for evaluating the safety of naturally driven vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of the virtual acceleration test method for an autonomous driving vehicle provided by the present invention.

[0062] Figure 2 This is a curve chart of high-precision fitting of natural driving data provided by the present invention.

[0063] Figure 3 The probability distribution diagram of the optimal importance sampling function provided by the present invention.

[0064] Figure 4 A comparison chart of scene conflict rate evaluation using different models provided by the present invention.

[0065] Figure 5 A comparison chart of the conflict rate convergence achieved by different models provided by the present invention. DETAILED DESCRIPTION

[0066] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0067] like Figure 1 As shown, the automatic driving vehicle virtual acceleration test method based on the optimal distribution model of the present invention comprises the following steps:

[0068] Step 1: Collect natural driving data, collect and record the vehicle's driving trajectory and motion status, including extracting the front vehicle's speed when the front vehicle cuts in V p , Test vehicle speed V c And the distance between the two cars R For example, the vehicle video or image information can be collected through the camera, the vehicle distance information can be collected through the lidar, and the vehicle location information can be obtained through GPS.

[0069] Step 2: Count key variables (spacing R ) to obtain the probability distribution of R The probability density histogram is then fitted with a mixed piecewise function to obtain the distribution characteristics of the key variables; the mixed piecewise function includes the exponential distribution function, the normal distribution function and the Pareto distribution function.

[0070] Step 3: Use the optimal importance sampling function to change the original distribution characteristics of key variables, increase the sampling probability of high-risk events, take the minimum number of tests as the optimization goal, construct the maximum likelihood estimation function, and obtain the optimal distribution for accelerated testing of high-risk events.

[0071] Step 4: Reconstruct high-risk test scenarios, conduct safety tests on autonomous vehicles in a simulation test platform, and calculate the probability of high-risk events and the acceleration ratio of the test.

[0072] Step 5: Evaluate test efficiency.

[0073] The automatic driving vehicle virtual acceleration test system based on the optimal distribution model of the present invention includes a natural driving data acquisition module, a key variable distribution fitting module, an optimal importance sampling module and a virtual scene simulation test module.

[0074] The above-mentioned natural driving data acquisition module provides data collection for functional scenarios, logical scenarios and specific scenarios for specific test requirements of the autonomous driving system. It collects vehicle video or image information through cameras, collects vehicle distance information through lidar, and obtains vehicle location information through GPS; the key variable distribution fitting module performs probability density statistics on key variables that can evaluate the danger of the scene, and selects appropriate distribution functions to fit their probability density histograms; the optimal importance sampling module evaluates the test efficiency of the original three distribution functions, and selects the distribution function with higher probability of danger and the least number of tests to replace the original distribution function; the virtual scene simulation test module resamples the optimal importance sampling function in the optimal importance sampling module, generates a new test scene, and evaluates the safety performance of the test vehicle in the simulation platform. Among them, the key variable distribution fitting module counts the front and rear distance R of the two vehicles as a probability density histogram and divides it into equal intervals, selects the fitting distribution function with the highest fitting accuracy in each small histogram segment as the best fitting function of the small histogram interval, and the best fitting function is selected from the exponential distribution function, the normal distribution function, and the Pareto distribution function.

[0075] The key variable distribution fitting module of this embodiment first collects the distance between the two vehicles before and after R The data is counted as a probability density histogram; then according to the spacing R The range of R The probability density histogram is divided into equally spaced segments, and the fitting distribution function with the highest fitting accuracy in each small segment of the histogram is selected as the best fitting function of the small segment of the histogram. The specific method for obtaining the best fitting function is as follows:

[0076] Step 1) Set X is a sequence of values ​​at the center of each probability histogram, Y For the sequence X The sequence of median values ​​and corresponding probability density values, is the probability distribution function in the sequence X The function value at each element in is:

[0077]

[0078]

[0079]

[0080] in, For thei The value of the center position of the probability histogram; for The corresponding probability density value; for The corresponding probability distribution function value; m is the number of bins contained in the histogram, ;

[0081] Step 2), Calculate Y and The absolute error is used to measure the fitting accuracy of each probability distribution function; the function with the highest fitting accuracy in each histogram interval should satisfy the following formula:

[0082]

[0083] In the formula, For the i The best fitting function of the segment; k are the numbers corresponding to the three probability distribution functions, 1, 2 and 3 correspond to the exponential distribution function, the normal distribution function and the Pareto distribution function respectively; For the k The probability distribution function is i The function value of the segment; for X A sequence of the best fitting function values ​​at each element in , where * indicates optimal.

[0084] After obtaining the best fitting function, the importance sampling method is then used to generate a probability distribution function with a higher probability of danger to replace the original probability distribution function. The maximum likelihood estimation function is first used. It represents the relationship between the optimal probability distribution function and the original probability distribution function, which is specifically expressed as follows:

[0085]

[0086] In the formula, It is the function with the highest fitting accuracy for natural driving data; is the optimal probability distribution function generated after importance sampling;

[0087] High risk areas are , is the indicator function of high risk areas, The probability of occurrence of high-risk events is:

[0088]

[0089]

[0090] In the formula, nis the number of tests; j For test samples;

[0091] To make exist Convergence is achieved at a confidence level of , the number of tests corresponding to the optimal distribution function n Should meet:

[0092]

[0093] In the formula, is the importance sampling parameter; is the likelihood ratio function; is the confidence level; is the cumulative distribution function value corresponding to the confidence level; is the convergence threshold.

[0094] In the process of actual measurement transformation of the original probability distribution function, it is necessary to improve the sampling probability of high-risk events while ensuring its unbiasedness; the measurement transformations corresponding to the original exponential distribution function, normal distribution function and Pareto distribution function are:

[0095]

[0096] k are the numbers corresponding to the three probability distribution functions, 1, 2 and 3 correspond to the exponential distribution function, the normal distribution function and the Pareto distribution function respectively; , are the coefficients of the three probability distribution functions respectively; is the corresponding acceleration parameter;

[0097] The unbiased constraint is that the original distribution function and the importance sampling function JS The divergence is less than or equal to 0.1, that is:

[0098] .

[0099] The virtual scene simulation test module of this embodiment includes a scene reconstruction module, a simulation test module and a simulation evaluation module. The scene reconstruction module is to resample the optimal importance sampling function to generate a new test scene. The simulation test module refers to building a virtual test platform for the front vehicle cutting-in scenario and using the test scenario as the initial state of the simulation test. The behavior of the test vehicle is controlled through the vehicle control model and the driving status of the vehicle is recorded. The simulation evaluation module refers to evaluating the danger probability of the scenario and the test mileage required for the danger probability to converge based on the collision rate of each test, thereby evaluating the test efficiency of the algorithm.

[0100] The test efficiency is calculated based on The number of tests and test mileage at which the probability of occurrence of high-risk scenarios reaches convergence at a confidence level;

[0101] The test mileage of the autonomous vehicle in natural driving scenarios and acceleration scenarios is compared using the acceleration ratio to evaluate the test efficiency of the hybrid segmented acceleration test method.

[0102] The calculation formula is as follows:

[0103]

[0104]

[0105]

[0106] In the formula, is the speedup ratio of the two test methods; Test mileage for autonomous vehicles in natural driving scenarios; Test mileage for autonomous vehicles in segmented acceleration scenarios; is the frequency of the front vehicle cutting in during natural driving scenarios; The number of tests of autonomous vehicles in natural driving scenarios; The number of tests for autonomous vehicles in mixed segment acceleration scenarios; The moment when the scene of the front car cuts in occurs; For the n The time the vehicle was driven during the test. Example

[0107] This embodiment collects the driving trajectory and motion state of vehicles in the highway scene, and extracts the speed of the front vehicle when the front vehicle cuts in. V p , Test vehicle speed V c And the distance between the two cars R . Effective scenarios meet , , Through data screening and processing, a total of 105,751 sets of valid scenes were extracted.

[0108] like Figure 2 As shown, statistics of each scene R The probability density histogram is obtained by fitting the exponential distribution function, normal distribution function, Pareto distribution function and other mixed distribution functions.

[0109] remember X and Yare sequences consisting of the values ​​at the center of each histogram and the corresponding probability density values, is the probability distribution function in the sequence X The function value at each element in is:

[0110]

[0111]

[0112]

[0113] in, For the i The value of the center position of the histogram; for The corresponding probability density value; for The corresponding probability distribution function value; m is the number of bins contained in the histogram, .

[0114] By calculation Y and The absolute error is used to measure the fitting accuracy of each probability distribution function, and the function with the highest fitting accuracy in each small segment of the histogram is selected as the best fitting function for the histogram interval. The function with the highest fitting accuracy in each histogram interval should satisfy the following formula:

[0115]

[0116] In the formula, For the i The best fitting function of the segment; k are the numbers corresponding to the three probability distribution functions, 1, 2 and 3 correspond to the exponential distribution function, the normal distribution function and the Pareto distribution function respectively; For the k The probability distribution function is i The function value of the segment; for X is a sequence of the best fitting function values ​​at each element in .

[0117] In order to increase the probability of dangerous events, this embodiment is as follows Figure 3 Importance sampling is used to generate a probability distribution function with a higher probability of danger to replace the original probability distribution function. The relationship between the optimal probability distribution function and the original probability distribution function can be expressed by the likelihood ratio L express:

[0118]

[0119] In the formula, It is the function with the highest fitting accuracy for natural driving data; is the optimal probability distribution function generated after importance sampling.

[0120] The relative distance between two vehicles is less than 150m, which is a high-risk area. , is the indicator function of high risk areas, The probability of a high-risk event occurring.

[0121]

[0122]

[0123] To make exist Convergence is achieved at a confidence level of , the number of tests corresponding to the optimal distribution function n Should meet:

[0124]

[0125] In the formula, is the importance sampling parameter.

[0126] The original probability distribution function is transformed by measurement to increase the sampling probability of high-risk events while ensuring its unbiasedness. The measurement transformations corresponding to the three original distribution functions are:

[0127]

[0128] The unbiased constraint is that the original distribution function and the importance sampling function JS The divergence is less than or equal to 0.1, that is:

[0129]

[0130] Build a virtual test platform for the front vehicle cut-in scenario, and test the vehicle equipped with an adaptive cruise control model and an automatic emergency braking model. Resample the optimal importance sampling function to generate a new test scenario , and the test scenario As the initial state of the simulation test, the behavior of the test vehicle is controlled through the vehicle control model, and the driving state of the vehicle, the probability of occurrence of dangerous events, and the number of tests are recorded.

[0131] Figure 4 and Figure 5 The test results of using the optimal distribution method in the technical solution of the present invention and using the existing natural driving model are shown.

[0132] from Figure 4It can be seen that the evaluation of the scene hazard rate by the present invention and the prior art is almost consistent.

[0133] from Figure 5 It can be seen that, under the condition of ensuring that the hazard rate remains unchanged, the present invention can achieve test convergence in 15877 times, while the prior art requires 3041480 times to achieve test convergence.

[0134] The speedup ratio is used to characterize the test efficiency of the model. The speedup ratio is calculated based on the number of tests and test mileage when the autonomous driving car reaches convergence in the natural driving scene and the acceleration scene test:

[0135]

[0136]

[0137]

[0138] Wherein, is the acceleration ratio of the two test methods; is the test mileage of the autonomous driving car in the natural driving scene; is the test mileage of the autonomous driving car in the segmented acceleration scene; is the frequency of the front vehicle cutting in the natural driving scene; is the number of tests of the autonomous driving car in the natural driving scene; is the number of tests of the autonomous driving car in the mixed segmented acceleration scene; is the time when the front vehicle cuts in the scene; is the first n The time the vehicle was driven during the test.

[0139] Calculation results show that the technical solution of the present invention can be accelerated by 560 times, significantly improving the efficiency of vehicle testing.

Claims

1. A virtual acceleration test method for an autonomous driving vehicle based on an optimal distribution model, characterized in that: The following steps are involved: Step 1: Collect natural driving data, collect and record the vehicle's driving trajectory and motion state, including extracting the front vehicle's speed V at the time of the front vehicle's cut-in p , the speed of the test vehicle V c And the distance R between the front and rear vehicles; Step 2: Calculate the probability distribution of the spacing R to obtain the probability density histogram of R in the scene, and then fit the probability density histogram using a mixed piecewise function to obtain the distribution characteristics of the key variables; the mixed piecewise function includes the exponential distribution function, the normal distribution function, and the Pareto distribution function; the specific method is: Then, the probability density histogram of R is divided into equally spaced segments according to the range of the spacing R, and the fitting distribution function with the highest fitting accuracy in each small segment of the histogram is selected as the best fitting function of the small segment of the histogram interval. The method for obtaining the best fitting function is as follows: Step 1), let X be the sequence of values ​​at the center of each probability histogram, and Y be the sequence of probability density values ​​corresponding to the median value of sequence X, is the function value of the probability distribution function at each element in the sequence X, that is: X=[x1,x2,...x i ,…,x m ] Y=[y1,y2,…,y i ,…,y m ] Among them, x i is the value of the center position of the i-th probability histogram; y i For x i The corresponding probability density value; For x i The corresponding probability distribution function value; m is the number of bins contained in the histogram, i = 1, 2, ..., m; Step 2), by calculating Y and The absolute error is used to measure the fitting accuracy of each probability distribution function; the function with the highest fitting accuracy in each histogram interval should satisfy the following formula: In the formula, is the best fitting function for the i-th segment; k is the number corresponding to the three probability distribution functions, 1, 2 and 3 correspond to the exponential distribution function, the normal distribution function and the Pareto distribution function respectively; is the function value of the kth probability distribution function in the i-th segment; is a sequence of the best fitting function values ​​at each element in X, * indicates optimal; Step 3: Use the optimal importance sampling function to change the original distribution characteristics of key variables, increase the sampling probability of high-risk events, take the minimum number of tests as the optimization goal, construct the maximum likelihood estimation function, and obtain the optimal distribution for accelerated testing of high-risk events; when the step 3 uses the importance sampling method to generate a probability distribution function with a higher probability of danger to replace the original probability distribution function, it is necessary to first use the maximum likelihood estimation function L(x) to represent the relationship between the optimal probability distribution function and the original probability distribution function, which is specifically expressed as follows: In the formula, f * (x) is the function with the highest fitting accuracy for natural driving data; is the optimal probability distribution function generated after importance sampling; The high risk area is ω, I ε (x) is the indicator function of the high-risk area, and γ is the probability of occurrence of high-risk events: In the formula, n is the number of tests; j is the test sample; In order to make γ converge at a confidence level of 100(1-α)%, that is, the relative half-width is less than a constant β, the number of tests n corresponding to the optimal distribution function should satisfy: Where θ is the importance sampling parameter; L 2 (x) is the likelihood ratio function; α is the confidence level; z α is the cumulative distribution function value corresponding to the confidence level; β is the convergence threshold; Step 4: Reconstruct high-risk test scenarios, conduct safety tests on autonomous vehicles in the simulation test platform, and calculate the probability of high-risk events and the acceleration ratio of the test; Step 5: Evaluate test efficiency.

2. The method for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model according to claim 1, characterized in that: The step 1 collects vehicle video or image information through a camera, collects vehicle distance information through a lidar, and obtains vehicle location information through a GPS.

3. The method for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model according to claim 1, characterized in that: The original probability distribution function is transformed by importance sampling method to improve the sampling probability of high-risk events while ensuring its unbiasedness. The measurement transformations corresponding to the original exponential distribution function, normal distribution function, and Pareto distribution function in step 2 are: k is the number corresponding to the three probability distribution functions, 1, 2 and 3 correspond to the exponential distribution function, normal distribution function and Pareto distribution function respectively; a i 、b i are the coefficients of the three probability distribution functions respectively; θ i is the corresponding acceleration parameter; The constraint condition for unbiasedness is that the JS divergence between the original distribution function and the importance sampling function is less than or equal to 0.1, that is: 。 4. The method for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model according to claim 1, characterized in that: The detailed process of step 4 is as follows: Step 4.1: Scenario reconstruction, that is, resampling the optimal importance sampling function to generate a new test scenario After importance sampling, the new spacing distribution is obtained through the optimal distribution model; Step 4.2: Simulation test, that is, build a virtual test platform for the front vehicle cutting-in scenario and As the initial state of the simulation test; control the behavior of the test vehicle through the vehicle control model and record the driving state of the vehicle; Step 4.3: Simulation evaluation, that is, based on the collision rate of each test, evaluate the danger probability of the scenario and the test mileage required for the danger probability to converge, thereby evaluating the test efficiency.

5. According to the method for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model in claim 4, the test efficiency refers to the number of tests and test mileage calculated when the probability of occurrence of high-risk scenarios reaches convergence at a confidence level of 100(1-α)%; The test mileage of the autonomous vehicle in natural driving scenarios and acceleration scenarios is compared using the acceleration ratio to evaluate the test efficiency of the hybrid segmented acceleration test method. The calculation formula is as follows: D nature =r lc ·N nature In the formula, r acc is the speedup ratio of the two test methods; D nature The test mileage of the autonomous vehicle in natural driving scenarios; D acc The test mileage of the autonomous vehicle in the segmented acceleration scenario; lc is the frequency of the front vehicle cutting in in the natural driving scene; N nature N is the number of tests of the autonomous vehicle in natural driving scenarios; acc The number of tests of the autonomous vehicle in the mixed segment acceleration scenario; t LC is the moment when the front vehicle cuts into the scene; τ LC is the driving time of the vehicle in the nth test.

6. A system for implementing the method for virtual acceleration testing of an autonomous driving vehicle based on an optimal distribution model as described in any one of claims 1 to 5, characterized in that: It includes natural driving data acquisition module, key variable distribution fitting module, optimal importance sampling module and virtual scene simulation test module; The natural driving data acquisition module provides data acquisition for functional scenarios, logical scenarios and specific scenarios for specific testing requirements of the autonomous driving system, collects vehicle video or image information through a camera, collects vehicle distance information through a lidar, and obtains vehicle location information through GPS; The key variable distribution fitting module performs probability density statistics on key variables that can assess the danger of a scene, and selects a suitable distribution function to fit its probability density histogram; The optimal importance sampling module evaluates the test efficiency of the original three distribution functions, and selects a distribution function with a higher probability of danger and the least number of tests to replace the original distribution function; The virtual scene simulation test module resamples the optimal importance sampling function in the optimal importance sampling module to generate a new test scene, and evaluates the safety performance of the test vehicle in the simulation platform.

7. The automatic driving vehicle virtual acceleration test system based on the optimal distribution model according to claim 6, characterized in that: The key variable distribution fitting module counts the front-to-rear distance R between the two vehicles as a probability density histogram and divides it into equally spaced segments, and selects the fitting distribution function with the highest fitting accuracy in each small histogram segment as the best fitting function of the small histogram interval. The best fitting function is selected from the exponential distribution function, the normal distribution function, and the Pareto distribution function.

Citation Information

Cited By

  • Automatic driving acceleration test method compatible with variable correlation

    CN120086493A