Method and system for testing acceleration performance of unmanned motorcycle under extreme working condition

Abstract

The invention relates to a method and a system for testing acceleration performance of an unmanned motorcycle under an extreme working condition, and belongs to the technical field of unmanned motorcycle testing. The method specifically comprises the following steps: step 1, generating an overall scene library according to an open source scene data set generated by an automatic driving actual test; step 2, collecting a part of scenes of the unmanned motorcycle subjected to extreme working condition testing in a real environment as a part of extreme working condition scene data set, and training by using a generative adversarial learning method to obtain an extreme working condition virtual scene library which is based on the overall scene library and contains extreme working condition characteristics; 3, clustering is carried out by adopting unsupervised learning according to the generated extreme working condition virtual scene library, and random generation is carried out according to the proportion of various scenes obtained through clustering when an extreme working condition test is carried out according to the proportion; and 4, according to a test result of the unmanned motorcycle in the generated extreme working condition virtual scene library, carrying out acceleration performance test on the unmanned motorcycle.

Description

Technical Field

[0001] The invention relates to an acceleration performance test method and system of an unmanned motorcycle under extreme working conditions, and belongs to the technical field of unmanned motorcycle testing. Background Art

[0002] Compared with double-track vehicles, unmanned motorcycles, as a single-track unmanned vehicle, have the characteristics of good concealment, flexibility, and strong environmental adaptability. They have great application potential in three-dimensional terrain, material transportation in narrow spaces, environmental monitoring, military reconnaissance, etc. However, two-wheeled motorcycles are under-driven systems and have inherently unstable characteristics. Especially under extreme working conditions, the stability of unmanned motorcycles is greatly challenged. Only by conducting multiple non-repetitive extreme working condition tests can we ensure that they can maintain their performance under actual extreme working conditions.

[0003] However, the unmanned motorcycle itself is composed of a strongly coupled system of vehicle-environment-task. The test content includes modules and functions such as the unmanned motorcycle perception system, map and positioning, decision-making planning system and control execution system. It is necessary not only to test the above modules in terms of software functions, but also to test the overall performance in a real road environment. Therefore, it is necessary to design very complex scenarios and tasks to prove the reliability and effectiveness of the unmanned motorcycle. If we rely entirely on road testing, it will not only be inefficient, but also bring certain road safety problems. Therefore, scenario-based virtual simulation testing has become an important idea to solve the problem of unmanned vehicle performance verification. In terms of virtual simulation testing methods, the current mainstream methods are test matrix method, worst case method, Monte Carlo method and accelerated testing method.

[0004] The test matrix method predefines the test scenarios, and the vehicle passes the test and is evaluated subjectively and objectively. However, it is difficult to enumerate all the test scenarios. The good performance of the vehicle in the scenarios pre-generated by the test matrix method does not mean that it can continue to integrate good performance in untested environments; the worst case method is used to evaluate the driving performance of the vehicle under extreme conditions, by selecting the most extreme or severe occasions for analysis and evaluation. However, the worst case method does not take into account the probability of extreme situations, thus ignoring various other scenarios; the Monte Carlo method uses statistical models to fit natural driving data, and uses Monte Carlo simulation to test and evaluate the vehicle. However, due to the small number of dangerous scenarios and extreme conditions in natural driving data, the efficiency is low.

[0005] The above three methods all have their own shortcomings, so we consider using accelerated testing methods. By constructing a virtual test environment, autonomous driving vehicles can be tested in this test environment just like on a real public road, but with higher efficiency. At the same time, the problem of too few valid samples in the sparsity disaster is solved. Summary of the invention

[0006] In view of this, the present invention proposes an acceleration performance test method and system for an unmanned motorcycle under extreme working conditions, which focuses on three aspects: construction of an overall scenario library, enhanced generation of a specific test scenario library for extreme working conditions, and multi-dimensional evaluation of test results.

[0007] The technical solution for implementing the present invention is as follows:

[0008] In the first aspect, the acceleration performance test method of the unmanned motorcycle under extreme working conditions of the present invention comprises the following specific processes:

[0009] Step 1: Generate an overall scenario library based on the open source scenario dataset generated by actual autonomous driving tests;

[0010] Step 2: Collect some scenes of unmanned motorcycles undergoing extreme working condition tests in real environments as some extreme working condition scene data sets, use the generative adversarial learning method, take the overall scene library as the original input, and take some extreme working condition sets as the target, and conduct adversarial training on the two to obtain an extreme working condition virtual scene library that is based on the overall scene library and contains extreme working condition characteristics;

[0011] Step 3: clustering the generated extreme working condition virtual scene library using unsupervised learning, and randomly generate scenes according to the proportions of various scenes obtained by clustering when conducting extreme working condition tests;

[0012] Step 4: Perform an acceleration performance test on the unmanned motorcycle based on the test results of the unmanned motorcycle in the generated extreme working condition virtual scene library.

[0013] Furthermore, the specific process of step one of the present invention is:

[0014] (1) Merge typical autonomous driving scene datasets and edge scene datasets into an overall scene dataset;

[0015] (2) According to the similarity with the extreme working condition scenario, fuzzy rules are set, the membership function is designed, the membership of each scenario in the overall data is calculated, and the scenarios with high membership are extracted as the initial overall scenario set; wherein the membership function is:

[0016] S = αF(e wather )+βG(e road )+γH(e object )

[0017] Among them, F(e weather )、G(e road )、H(e object ) are the membership functions based on weather feature deviation, road feature deviation and surrounding object feature deviation respectively, α, β and γ are weighting coefficients, and α+β+γ=1.

[0018] Furthermore, the present invention finally screens out scenes with a membership degree greater than 50% as the overall scene data set.

[0019] Furthermore, after obtaining the extreme working condition virtual scene library, step 2 of the present invention also transforms and combines the scenes based on the specific scenarios in the overall scene library and the characteristics of the actual extreme working condition scenes to generate different extreme working condition test samples.

[0020] Furthermore, the specific process of using the generative adversarial learning method in step 2 of the present invention is:

[0021] A part of the scenes of unmanned motorcycles under real extreme working conditions is used as real data, that is, learning objects. The training method of generative adversarial network is adopted. The initial generator is initialized with the overall scene set, and random noise is added as the generated data. During training, first, the generator is fixed, and the discriminator is trained to accurately distinguish between real samples and fake samples; secondly, the discriminator is fixed, and the generator is trained to generate samples that are sufficient to deceive the discriminator.

[0022] Furthermore, the adversarial network value function using the generative adversarial learning method in step 2 of the present invention is as follows:

[0023]

[0024] Among them, P data (x) is the real data distribution, P z (z) is the generated data distribution, z is the noise, D(x) is the discriminant result of the real data, and D(G(z))) is the discriminant result of the generated data.

[0025] Furthermore, the cost functions of the discriminator and the generator of the present invention are respectively:

[0026]

[0027] For the discriminator, the larger V(G,D) is, the better; for the generator, the smaller V(G,D) is, the better.

[0028] Furthermore, in the present invention, the unsupervised clustering of scene data in step three adopts a K-means clustering method, and for a given scene set, the samples are divided into K clusters according to the distances between the samples.

[0029] Furthermore, the present invention calculates the distance between numerical objects by Euclidean distance.

[0030]

[0031] Where k is the number of clusters, C iis the point set of the i-th cluster, x is the point belonging to C i Data points, μ i is the center of the ith cluster.

[0032] The selection should satisfy E>E min The test sample, E is the distance between the samples obtained by clustering, E min The minimum sample distance to be set.

[0033] In a second aspect, the acceleration performance test system of an unmanned motorcycle under extreme working conditions of the present invention comprises: a data acquisition module, a data generation module, a clustering module and an acceleration performance test module;

[0034] The data collection module is used to collect and generate an overall scenario library based on the open source scenario data set generated by actual autonomous driving tests;

[0035] The data generation module is used to collect some scenes of unmanned motorcycles undergoing extreme working condition tests in real environments as some extreme working condition scene data sets. The generative adversarial learning method is used to take the overall scene library as the original input and the partial extreme working condition set as the target. The two are trained adversarially to obtain an extreme working condition virtual scene library that is based on the overall scene library and contains extreme working condition characteristics.

[0036] The data clustering module is used to cluster the generated extreme working condition virtual scene library using unsupervised learning, and randomly generate scenes in proportion according to the proportions of various scenes obtained by clustering when conducting extreme working condition tests;

[0037] The acceleration performance test module is used to perform acceleration tests on the unmanned motorcycle based on the test results of the unmanned motorcycle in the generated extreme working condition virtual scene library.

[0038] Beneficial effects:

[0039] The present invention uses adversarial learning as the main method, extracts traditional scenarios and trains a generator based on adversarial learning, so as to generate the required extreme operating conditions. Then, based on clustering and according to the probability of occurrence, the unmanned motorcycle is placed in extreme operating conditions that appear in proportion, and the distance between test scenarios is limited to avoid repeated testing as much as possible. Finally, the key indicators of extreme operating conditions are evaluated according to various categories and multi-dimensional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 Flow chart of the performance test method of the present invention for unmanned motorcycle under extreme working conditions

[0042] Figure 2 is a flow chart of the adversarial training method of the present invention;

[0043] Figure 3 This is the clustering result of the present invention. DETAILED DESCRIPTION

[0044] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0045] It should be noted that the following embodiments and features in the embodiments may be combined with each other in the absence of conflict; and, based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making any creative work are within the scope of protection of the present disclosure.

[0046] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0047] The present application embodiment provides a method for testing the acceleration performance of an unmanned motorcycle under extreme working conditions, such as Figure 1 As shown, the specific process is:

[0048] Step 1: Generate an overall scenario library based on the open source scenario dataset generated by actual autonomous driving tests.

[0049] Before the overall scenario library is generated, this step also sets fuzzy rules according to the possible scenarios of the unmanned motorcycle's extreme working conditions, eliminates scenes with low membership, and retains scenes with higher membership.

[0050] When this step is implemented, follow the following process:

[0051] (1) Overall scene data acquisition: The typical scene data sets and edge scene data sets of autonomous driving are merged into an overall scene data set.

[0052] (2) According to the similarity with the extreme working condition scenario, fuzzy rules are set, the membership function is designed, the membership of each scenario in the overall data is calculated, and the scenarios with high membership are extracted as the initial overall scenario set; wherein the membership function is:

[0053] S = αF(e weather )+βG(e road )+γH(e object )

[0054] Among them, F(e weather )、G(e road )、H( eobject ) are the membership functions based on weather feature deviation, road feature deviation and surrounding object feature deviation respectively, α, β, γ are weighting coefficients, α+β+γ=1. Finally, the scenes with a membership greater than 50% are selected as the overall scene data set.

[0055] Step 2: Collect some scenes of unmanned motorcycles undergoing extreme working condition tests in real environments as some extreme working condition scene data sets, use the generative adversarial learning method, take the overall scene library as the original input, and take some extreme working condition sets as the target, and perform adversarial training on the two to obtain an extreme working condition virtual scene library that is based on the overall scene library and contains extreme working condition characteristics.

[0056] After obtaining the extreme working condition virtual scene library, this step also transforms and combines the scenes based on the specific scenarios in the overall scene library and the characteristics of the actual extreme working condition scenes, so as to generate enough different extreme working condition test samples.

[0057] In the specific implementation of this step, adversarial learning is performed according to the following process: Figure 2 As shown:

[0058] A portion of the scenes of unmanned motorcycles under real extreme working conditions are used as real data, that is, learning objects. The training method of the generative adversarial network (including generator and discriminator) is adopted. The initial generator is initialized with the overall scene set, and random noise is added as the generated data. First, the generator is fixed, hoping that the discriminator can accurately distinguish between real samples and fake samples; second, the discriminator is fixed, hoping that the generator can generate samples that are sufficient to deceive the discriminator. The overall value function is as follows

[0059]

[0060] Among them, P data (x) is the real data distribution, D z (z) is the generated data distribution, z is the noise, D(x) is the discriminator's result on the real data, and D(G((z))) is the discriminator's result on the generated data;

[0061] The cost functions of the discriminator and generator are:

[0062]

[0063] For the discriminator D, the larger V(G,D) is, the better; and for the generator G, the smaller V(G,D) is, the better. During training, always fix one of them and train the other.

[0064] The ultimate goal is to train a generator that is good enough to randomly generate a library of virtual scenarios for extreme working conditions.

[0065] Step 3: cluster the generated extreme working condition virtual scene library using unsupervised learning to generate several different extreme working condition scenes. According to the proportion of various scenes obtained by clustering, they are randomly generated according to the proportion when conducting extreme working condition tests.

[0066] In this step, a generator is used to randomly generate scenario data, and unsupervised clustering is further used to derive scenarios with various characteristics, proportionally enhancing the probability of occurrence of high-probability extreme conditions to avoid serious deviations in test results caused by bad data and duplicate data, thereby affecting the control optimization direction of the unmanned motorcycle.

[0067] In the specific implementation of this step, the K-means clustering method is used for the unsupervised clustering of scene data. For a given scene set, the samples are divided into K clusters according to the distance between the samples. K-means clustering calculates the distance between numerical objects through Euclidean distance.

[0068]

[0069] Where k is the number of clusters, C i is the point set of the i-th cluster, x is the point belonging to C i Data points, μ i is the center of the ith cluster, and represents the square of the Euclidean distance between x and the cluster center. K-means clustering makes the points in the cluster as close together as possible and the distance between clusters as large as possible. The clustering results are as follows: Figure 3 shown.

[0070] Through the K-means clustering method, we can get several virtual extreme scenarios with different numbers of sample scenarios, which are divided into high-probability scenarios and low-probability scenarios according to the number of samples. At the same time, considering that the generated sample library inevitably has a certain number of repeated samples, the distance between the numerical objects recorded in the clustering process should not be too small to avoid repeated results in the test process as much as possible, and try to accelerate the test. The selection of test samples should meet the following requirements:

[0071] E>E min

[0072] The test samples are given by K-means clustering in proportion, and E is the distance between each sample obtained by clustering. min The minimum sample distance to set to avoid repeated testing as much as possible.

[0073] Step 4: According to the test results of the unmanned motorcycle in the generated extreme working condition virtual scene library, the acceleration performance of the unmanned motorcycle is tested and evaluated.

[0074] Performance testing and evaluation in this step: Evaluation is made based on the response data of the unmanned motorcycle during the virtual scene test as key performance indicators. The key performance under extreme conditions includes the stability of the unmanned motorcycle, using the lateral speed, roll angle and roll angle speed of the unmanned motorcycle during the test as evaluation criteria; the passability of the unmanned motorcycle is evaluated based on the width and height of the obstacles that can be crossed.

[0075] For example, the stability of the unmanned motorcycle during movement is measured based on its roll angle, roll angular velocity and lateral acceleration; the anti-interference ability of the unmanned motorcycle's self-balancing control system is evaluated by indicators such as the roll angular velocity and lateral acceleration speed after a horizontal collision.

[0076] The present invention uses a reinforcement learning method to learn the overall scenario library, and conducts adversarial learning with the collected specific scenario data of some extreme conditions, thereby generating a virtual simulated extreme condition scenario library. The unmanned motorcycle is tested under the extreme condition virtual scenario library generated by adversarial learning. The extreme condition virtual scenario library makes up for the shortcomings of the corresponding extreme condition scenarios in the real scene. At the same time, the extreme condition test in the virtual scene also meets the goal of accelerated testing.

[0077] In addition, the present invention clusters the generated virtual scene library and performs tests according to the occurrence ratio, thereby simulating the occurrence of various types of extreme working conditions of the unmanned motorcycle in a real environment, avoiding the problems of large-scale testing of low-probability extreme working conditions and low-probability testing under high-probability extreme working conditions, thereby achieving accelerated testing.

[0078] The embodiment of the present application provides an acceleration performance test system for an unmanned motorcycle under extreme working conditions, comprising: a data acquisition module, a data generation module, a clustering module and an acceleration performance test module;

[0079] The data collection module is used to collect and generate an overall scenario library based on the open source scenario data set generated by actual autonomous driving tests;

[0080] The data generation module is used to collect some scenes of unmanned motorcycles undergoing extreme working condition tests in real environments as some extreme working condition scene data sets. The generative adversarial learning method is used to take the overall scene library as the original input and the partial extreme working condition set as the target. The two are trained adversarially to obtain an extreme working condition virtual scene library that is based on the overall scene library and contains extreme working condition characteristics.

[0081] The data clustering module is used to cluster the generated extreme working condition virtual scene library using unsupervised learning, and randomly generate scenes in proportion according to the proportions of various scenes obtained by clustering when conducting extreme working condition tests;

[0082] The acceleration performance test module is used to perform acceleration tests on the unmanned motorcycle based on the test results of the unmanned motorcycle in the generated extreme working condition virtual scene library.

[0083] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. The acceleration performance test method of an unmanned motorcycle under extreme working conditions is characterized by: The specific process is: Step 1: Generate an overall scenario library based on the open source scenario dataset generated by actual autonomous driving tests; Step 2: Collect some scenes of unmanned motorcycles undergoing extreme working condition tests in real environments as some extreme working condition scene data sets, use the generative adversarial learning method, take the overall scene library as the original input, and take some extreme working condition sets as the target, and conduct adversarial training on the two to obtain an extreme working condition virtual scene library that is based on the overall scene library and contains extreme working condition characteristics; Step 3: clustering the generated extreme working condition virtual scene library using unsupervised learning, and randomly generate scenes according to the proportions of various scenes obtained by clustering when conducting extreme working condition tests; Step 4: Perform an acceleration performance test on the unmanned motorcycle based on the test results of the unmanned motorcycle in the generated extreme working condition virtual scene library.

2. The acceleration performance test method of an unmanned motorcycle under extreme working conditions according to claim 1, characterized in that: The specific process of step one is: (1) Merge typical autonomous driving scene datasets and edge scene datasets into an overall scene dataset; (2) According to the similarity with the extreme working condition scenario, fuzzy rules are set, the membership function is designed, the membership of each scenario in the overall data is calculated, and the scenarios with high membership are extracted as the initial overall scenario set; wherein the membership function is: S=αF(e weather )+βG(e road )+γH(e object ) Among them, F(e weather )、G(e road )、H(e object ) are the membership functions based on weather feature deviation, road feature deviation and surrounding object feature deviation respectively, α, β and γ are weighting coefficients, and α+β+γ=1.

3. The acceleration performance test method under extreme working conditions of an unmanned motorcycle according to claim 2, characterized in that: The scenes with membership greater than 50% are finally screened out as the overall scene data set.

4. The acceleration performance test method of an unmanned motorcycle under extreme working conditions according to claim 1, characterized in that: After obtaining the extreme working condition virtual scene library, the step 2 is to transform and combine the actual extreme working condition scenes based on the specific scenarios in the overall scene library according to the characteristics of the actual extreme working condition scenes to generate different extreme working condition test samples.

5. The acceleration performance test method of an unmanned motorcycle under extreme working conditions according to claim 1, characterized in that: The specific process of using the generative adversarial learning method in step 2 is: A part of the scenes of unmanned motorcycles under real extreme working conditions is used as real data, that is, learning objects. The training method of generative adversarial network is adopted. The initial generator is initialized with the overall scene set, and random noise is added as the generated data. During training, first, the generator is fixed, and the discriminator is trained to accurately distinguish between real samples and fake samples; secondly, the discriminator is fixed, and the generator is trained to generate samples that are sufficient to deceive the discriminator.

6. The acceleration performance test method of an unmanned motorcycle under extreme working conditions according to claim 5 is characterized in that: The adversarial network value function used in step 2 is as follows: Among them, P data (x) is the real data distribution, P z (z) is the generated data distribution, z is the noise, D(x) is the discriminator’s judgment result on the real data, and D(G((z))) is the discriminator’s judgment result on the generated data.

7. The acceleration performance test method of an unmanned motorcycle under extreme working conditions according to claim 6 is characterized in that: The cost functions of the discriminator and generator are: For the discriminator, the larger V(G, D) is, the better; for the generator, the smaller V(G, D) is, the better.

8. The acceleration performance test method of an unmanned motorcycle under extreme working conditions according to claim 1 is characterized in that: In the step three, the K-means clustering method is used for the unsupervised clustering of the scene data. For a given scene set, the samples are divided into K clusters according to the distance between the samples.

9. The acceleration performance test method of an unmanned motorcycle under extreme working conditions according to claim 8, characterized in that: The distance between samples is calculated by Euclidean distance. Where k is the number of clusters, C i is the point set of the i-th cluster, x is the point belonging to C i Data points, μ i is the center of the i-th cluster; The selection should satisfy E>E min The test sample, E is the distance between the samples obtained by clustering, E min The minimum sample distance to be set.

10. The acceleration performance test system of unmanned motorcycle under extreme working conditions is characterized by: include: Data collection module, data generation module, clustering module and performance testing module; The data collection module is used to collect and generate an overall scenario library based on the open source scenario data set generated by actual autonomous driving tests; The data generation module is used to collect some scenes of unmanned motorcycles undergoing extreme working condition tests in real environments as some extreme working condition scene data sets. The generative adversarial learning method is used to take the overall scene library as the original input and the partial extreme working condition set as the target. The two are trained adversarially to obtain an extreme working condition virtual scene library that is based on the overall scene library and contains extreme working condition characteristics. The data clustering module is used to cluster the generated extreme working condition virtual scene library using unsupervised learning, and randomly generate scenes in proportion according to the proportions of various scenes obtained by clustering when conducting extreme working condition tests; The acceleration performance test module is used to perform acceleration tests on the unmanned motorcycle based on the test results of the unmanned motorcycle in the generated extreme working condition virtual scene library.

Images