Visual test system and method

Abstract

The invention discloses a visual testing system and method, and the system comprises a dynamic scene simulation module which is used for generating a dynamic testing scene through a Unity 3D or Unreal physical engine, and the adjustable parameters of the dynamic testing scene comprise the illumination intensity, the weather condition and the dynamic object density; the adversarial sample generation module is used for adding adversarial disturbance based on an FGSM or PGD algorithm into the input image by generating an adversarial network model or a diffusion model; and the sensor simulation module is used for simulating hardware output data of the camera, the inertial measurement unit or the laser radar, and the hardware output data comprises frame rate, noise and drift parameters. According to the invention, through cooperation of the dynamic scene parameter programmable control module and the anti-disturbance coupling injection module, multi-modal parameter nonlinear coupling adjustment of illumination intensity, object motion trail and anti-noise is realized, and the problem of test scene distortion caused by independent control of environmental parameters in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving testing, and in particular to a vision testing system and method. Background Art

[0002] With the commercialization of autonomous driving systems above Level 3, visual positioning algorithms need to maintain centimeter-level positioning accuracy in complex scenarios such as extreme weather and interference from dynamic obstacles. The existing ISO 34502 standard requires that the test system cover more than 99% of scenario parameter combinations, while traditional static test platforms can only simulate limited lighting and weather conditions and cannot construct continuously changing adversarial environments. Especially on urban roads, the coupling effect of sudden obstacles (such as pedestrians intruding, vehicles cutting in) and sensor noise (rain and snow reflections, lens damage) makes it difficult for existing test methods to reproduce real-world long-tail risk scenarios.

[0003] While existing technologies can control single environmental parameters (such as light intensity or obstacle density) through preset scripts, they cannot achieve coordinated dynamic adjustment of multimodal parameters (physical environment + countermeasures + sensor noise). For example, when simulating a rainstorm, existing systems can only independently adjust rainfall parameters and camera noise parameters. However, in real-world scenarios, the visual impact of raindrops is nonlinearly coupled with road reflections and reduced visibility. This parameter decoupling control method results in insufficient test coverage, and actual measurements show that its coverage of extreme scenarios is less than 60% of the standard requirement. Summary of the Invention

[0004] The object of the present invention is to provide a vision testing system and method to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A visual testing system comprising

[0007] A dynamic scene simulation module is used to generate a dynamic test scene using Unity3D or Unreal physics engine. The adjustable parameters of the dynamic test scene include light intensity, weather conditions, and dynamic object density.

[0008] Adversarial sample generation module, which is used to add adversarial perturbations based on FGSM or PGD algorithms to the input image by generating adversarial network models or diffusion models;

[0009] A sensor simulation module is used to simulate the hardware output data of a camera, inertial measurement unit, or lidar, wherein the hardware output data includes frame rate, noise, and drift parameters;

[0010] The test execution module is used to run the target visual positioning algorithm, adjust the test scene parameters according to the reinforcement learning strategy, and send the parameter adjustment records and algorithm output to the evaluation module;

[0011] The evaluation module is used to adjust the records and algorithm outputs according to the received parameters, calculate the positioning error, fault recovery time and energy consumption indicators, and generate a visual evaluation report.

[0012] In the present invention, the dynamic scene simulation module also includes a physical parameter adjustment function of wind speed and friction coefficient.

[0013] In the present invention, the extreme scene images synthesized by the adversarial sample generation module include foggy, snowy or dynamic occluder images.

[0014] In the present invention, the test execution module is used to record the posture output data of the target visual positioning algorithm during the test process. When the posture error exceeds a preset threshold, the current scene parameter combination is marked as a failure case. According to the frequency distribution of the failure case, the test weight of the corresponding scene parameter is dynamically adjusted.

[0015] A visual testing method comprises the following steps:

[0016] Step S1: parsing the input business scenario description through the natural language processing module to extract scenario keywords;

[0017] Step S2, generating a dynamic test scene corresponding to the scene keyword based on the physical engine, wherein the parameters of the dynamic test scene include light intensity and dynamic object density;

[0018] Step S3, adding adversarial perturbations to the dynamic test scene image through a generative adversarial network model, or synthesizing extreme weather images through a diffusion model;

[0019] Step S4: Run the target visual positioning algorithm, adjust the test scene parameters through the reinforcement learning agent, and record the parameter combination when the algorithm fails;

[0020] Step S5: Generate an evaluation report based on the posture error, fault recovery time, and energy consumption indicators, wherein the evaluation report includes a parameter sensitivity heat map.

[0021] It is characterized in that: in the step S3, the method of adding the adversarial perturbation is the fast gradient sign method or the projected gradient descent method.

[0022] It is characterized in that: in step S4, the reinforcement learning agent updates the probability distribution of the scene parameters through gradient back propagation.

[0023] It is characterized in that: the step S5 also includes directly applying the algorithm model under the training scene A to the test scene B, and when the positioning error of scene B exceeds 150% of the error of scene A, it is determined that the generalization is insufficient.

[0024] A computer-readable storage medium stores a computer program, which implements the method steps when executed by a processor.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The present invention realizes the nonlinear coupling adjustment of multimodal parameters such as light intensity, object motion trajectory, and anti-noise through the coordinated cooperation of the dynamic scene parameter programmable control module and the anti-disturbance coupling injection module, thus solving the test scene distortion problem caused by the independent control of environmental parameters in the prior art.

[0027] 2. The present invention realizes online adaptive adjustment of test parameters by strengthening the closed-loop feedback mechanism of the learning strategy optimization layer and the hardware-in-the-loop simulation architecture, overcoming the defect that the preset script adjustment method cannot respond to the real-time status of the system under test. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a diagram of the multimodal collaborative testing architecture of the present invention;

[0029] Figure 2 This is a flow chart of the adaptive boundary test of the present invention;

[0030] Figure 3 This is a timing diagram of the extreme weather test of the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] Example 1

[0033] This invention addresses the issue of insufficient test coverage for autonomous driving visual positioning systems in complex dynamic scenarios. It achieves high-fidelity, adaptive boundary testing through dynamic adversarial environment simulation and reinforcement learning-driven parameter optimization. Because it involves physical engine rendering, adversarial sample generation, and hardware-in-the-loop simulation, it relies on hardware architecture to achieve the following technical goals:

[0034] Programmable control of dynamic scene parameters (real-time adjustment of lighting / object density / weather);

[0035] Counteract the coupled injection of disturbance and sensor noise;

[0036] Online reinforcement learning optimization of testing strategies.

[0037] Based on the above technical requirements, this embodiment adopts Figure 1 The "multimodal collaborative testing architecture" shown:

[0038] Physical rendering layer: The motion trajectory of dynamic objects is simulated through the Niagara particle system of Unreal Engine 5. Its kinematic model is:

[0039]

[0040] in, is the initial velocity, Expressed as acceleration.

[0041] Adversarial attack layer: The PGD attacker and the diffusion model form a two-way perturbation generation pipeline. The PGD perturbation generation formula is:

[0042]

[0043] Among them, ∈ represents the perturbation boundary, α represents the step size, and J represents the loss function.

[0044] Strategy optimization layer: The PPO algorithm is used to adjust the test parameters. The strategy network output parameter adjustment amount Δp is:

[0045] Δp=μ(s|θ μ )+N t ·σ

[0046] Where s=[e pos ,t latency ,P gpu ],N t is denoted as exploration noise.

[0047] The dynamic scene simulation module and the adversarial sample generation module are interconnected via a PCIe 4.0x8 bus, with a data transmission bandwidth of ≥16GB / s;

[0048] The noise injection unit of the sensor simulation module is implemented using FPGA, and the delay jitter is controlled within ±2μs;

[0049] The error calculator of the evaluation module and the reinforcement learning agent share the DDR4 memory pool and achieve data synchronization through atomic locks.

[0050] like Figure 2 As shown, the adaptive boundary testing process performed on this architecture includes:

[0051] Step S1: Initialize dynamic scene parameters:

[0052] Parse the initial parameter value according to the input requirements:

[0053]

[0054] Step S2: Coupled injection countermeasure disturbance:

[0055] Apply illumination interference and PGD perturbation to the camera input image I:

[0056]

[0057] Among them, ΔI light Obeys the Rayleigh distribution R (σ=0.2).

[0058] Step S3: Reinforcement learning strategy update:

[0059] After each round of testing, update the policy network parameters θ:

[0060]

[0061] Among them, the advantage function A t The GAR method is used for calculation.

[0062] Step S4: Cross-modal evaluation:

[0063] Calculate the comprehensive test coverage C:

[0064]

[0065] Among them, e th is the error threshold, and Ⅱ is the indicative function.

[0066] Example 2

[0067] Suppose that a Level 4 autonomous driving system needs to pass a visual positioning test in extreme weather conditions.

[0068] like Figure 3 The implementation process is as follows:

[0069] Step 1: Initial parameter setting:

[0070] Training scene A is set to sunny day (light intensity L A =0.1Lux), test scene B simulates blizzard weather (L B =0.05Lux, snowflake density ρ snow =2000flakes / m 3 ).

[0071] Step 2: Generate adversarial disturbances:

[0072] Apply a physically based snow accumulation effect to the input image of scene B:

[0073]

[0074] Among them, α i is the transparency of the i-th snowflake, M i is the mask matrix.

[0075] Step 3: Assessment of forward mobility:

[0076] Calculate the performance degradation rate in the Blizzard scenario:

[0077]

[0078] When R decay When it is >150%, the parameter reconfiguration process is triggered.

[0079] Step 4, strategy optimization results:

[0080] After 3 rounds of PPO iterations, the optimal test parameters converge to:

[0081] p * ={light=0.08Lux,obj density =12 units / 100 m 2 ,∈PGD=0.04}

[0082] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0083] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A visual testing system, characterized in that: include A dynamic scene simulation module is used to generate a dynamic test scene using Unity3D or Unreal physics engine. The adjustable parameters of the dynamic test scene include light intensity, weather conditions, and dynamic object density. Adversarial sample generation module, which is used to add adversarial perturbations based on FGSM or PGD algorithms to the input image by generating adversarial network models or diffusion models; A sensor simulation module is used to simulate the hardware output data of a camera, inertial measurement unit, or lidar, wherein the hardware output data includes frame rate, noise, and drift parameters; The test execution module is used to run the target visual positioning algorithm, adjust the test scene parameters according to the reinforcement learning strategy, and send the parameter adjustment records and algorithm output to the evaluation module; The evaluation module is used to adjust the records and algorithm outputs according to the received parameters, calculate the positioning error, fault recovery time and energy consumption indicators, and generate a visual evaluation report.

2. A visual testing system according to claim 1, characterized in that: The dynamic scene simulation module also includes a physical parameter adjustment function for wind speed and friction coefficient.

3. A visual testing system according to claim 1, characterized in that: The extreme scene images synthesized by the adversarial sample generation module include foggy, snowy or dynamic occluder images.

4. A visual testing system according to claim 1, characterized in that: The test execution module is used to record the posture output data of the target visual positioning algorithm during the test process. When the posture error exceeds a preset threshold, the current scene parameter combination is marked as a failure case. According to the frequency distribution of the failure case, the test weight of the corresponding scene parameter is dynamically adjusted.

5. A visual testing method, characterized in that: The following steps are involved: Step S1: parsing the input business scenario description through the natural language processing module to extract scenario keywords; Step S2, generating a dynamic test scene corresponding to the scene keyword based on the physical engine, wherein the parameters of the dynamic test scene include light intensity and dynamic object density; Step S3, adding adversarial perturbations to the dynamic test scene image through a generative adversarial network model, or synthesizing extreme weather images through a diffusion model; Step S4: Run the target visual positioning algorithm, adjust the test scene parameters through the reinforcement learning agent, and record the parameter combination when the algorithm fails; Step S5: Generate an evaluation report based on the posture error, fault recovery time, and energy consumption indicators, wherein the evaluation report includes a parameter sensitivity heat map.

6. A visual testing system according to claim 1, characterized in that: In step S3, the adversarial perturbation is added by using the fast gradient sign method or the projected gradient descent method.

7. A visual testing system according to claim 1, characterized in that: In step S4, the reinforcement learning agent updates the probability distribution of the scene parameters through gradient back propagation.

8. A visual testing system according to claim 1, characterized in that: The step S5 further includes directly applying the algorithm model under the training scene A to the test scene B. When the positioning error of scene B exceeds 150% of the error of scene A, it is determined that the generalization is insufficient.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method steps of any one of claims 5 to 8 are implemented.

Images