Intelligent driving vehicle perception evaluation method and system based on multi-dimensional sensing

Abstract

The application provides a kind of intelligent driving vehicle perception evaluation method and system based on multidimensional sensing, belongs to vehicle perception evaluation technical field.The method comprises: collecting multi-source heterogeneous sensing data, obtaining a unified multidimensional perception data set;Multi-level environmental disturbance model is constructed to generate multidimensional perception degradation scene, form multidimensional perception degradation trajectory;Construct a semantic structure atlas, and construct a credibility reconstruction evaluation model through a pre-set rule constraint;Multi-dimensional fusion evaluation function containing dynamic weight adjustment mechanism is constructed, and the evaluation result is obtained;Self-evolution test scene generation mechanism is constructed, and the dynamic generation and optimization of the antagonistic test scene are carried out;Closed-loop feedback mechanism is constructed, and the multi-level environmental disturbance model and the credibility reconstruction evaluation model are adaptively updated.The application realizes the spatiotemporal high-precision alignment of multidimensional perception data, continuously depicts the perception performance degradation process and identifies the robustness boundary, automatically constructs the long tail scene and realizes effective coverage.

Description

Technical Field

[0001] This invention relates to the field of vehicle perception and evaluation technology, and specifically to a method and system for intelligent driving vehicle perception and evaluation based on multi-dimensional sensing. Background Technology

[0002] With the rapid development of autonomous driving and intelligent connected vehicle technologies, high-end models are typically equipped with multi-source heterogeneous sensing systems, including cameras, millimeter-wave radar, and lidar, to perceive and understand the vehicle's surrounding environment. Current technologies usually involve fusing multi-source sensor data to perform tasks such as target detection, obstacle recognition, and trajectory prediction. Based on this, a perception performance evaluation system is built to assess the system's performance under different road and traffic conditions.

[0003] However, existing evaluation methods largely rely on pre-set test scenarios and static evaluation indicators, primarily evaluating perception systems based on detection accuracy or average error under fixed operating conditions. This makes it difficult to reflect the true characteristics of perception performance changing over time in complex and dynamic traffic environments. Furthermore, in the process of multi-source perception information fusion, existing methods often focus on semantic-level recognition results, lacking effective verification mechanisms for their consistency with the actual physical laws of vehicle motion. In addition, existing testing systems typically rely on manual design or fixed scenario libraries, making it difficult to cover extremely complex or long-tail traffic scenarios, and limiting the ability to generate and update test scenarios.

[0004] In existing technologies, the evaluation of perception in intelligent driving vehicles still suffers from the following shortcomings: First, existing evaluation methods are usually based on statistical analysis of discrete test points, which cannot continuously characterize the performance changes of the perception system under different complex environmental disturbances, resulting in an incomplete and imprecise assessment of the system's robustness and stability. Second, existing technologies mainly evaluate based on semantic-level detection and recognition results, without fully considering whether the target motion conforms to vehicle dynamics and traffic physics, which may lead to discrepancies between the evaluation results and real traffic behavior. Third, most existing test scenarios rely on manual construction or fixed databases, making it difficult to automatically generate targeted test scenarios based on the weaknesses of the perception system, resulting in insufficient adaptability of the test system and an inability to meet the continuous verification needs in highly complex driving environments. Summary of the Invention

[0005] The purpose of this invention is to provide a perception evaluation method and system for intelligent driving vehicles based on multi-dimensional sensing, so as to solve the technical problems of existing intelligent driving vehicle perception evaluation methods, such as difficulty in tracing performance evolution due to multi-discrete sampling, lack of physical consistency verification, and difficulty in adaptive generation to address shortcomings due to fixed scenarios.

[0006] To achieve the above objectives, this invention provides a perception evaluation method for intelligent driving vehicles based on multi-dimensional sensing. The method includes: collecting multi-source heterogeneous sensing data from multi-dimensional sensors deployed on the vehicle to be evaluated, performing time synchronization and spatial calibration to obtain a unified multi-dimensional perception data set; constructing a multi-level environmental disturbance model based on the multi-dimensional perception data set to generate multi-dimensional perception degradation scenarios, recording the output changes of the vehicle's perception evaluation system at different degradation levels to form a multi-dimensional perception degradation trajectory; and constructing a semantic structure graph based on the multi-dimensional perception degradation trajectory, and constraining the multi-dimensional perception data using preset rules. The perceptual degradation trajectory is inverted and verified to construct a credibility reconstruction evaluation model. Based on the multidimensional perceptual degradation trajectory and credibility reconstruction evaluation results, a multidimensional fusion evaluation function with a dynamic weight adjustment mechanism is constructed to obtain the evaluation results. Based on the evaluation results and the analysis results of the multidimensional perceptual degradation trajectory, a self-evolving test scenario generation mechanism is constructed to identify the weak areas of the vehicle's perception evaluation system and perform reverse scenario modeling, thereby dynamically generating and optimizing adversarial test scenarios. Finally, based on the evaluation results and the target historical data of the multidimensional perceptual degradation trajectory, a closed-loop feedback mechanism is constructed to adaptively update the multi-level environmental disturbance model and the credibility reconstruction evaluation model.

[0007] Optionally, obtaining unified multidimensional sensing data includes: performing time-base unification processing on multi-source heterogeneous sensing data, aligning preset sensor data to a preset time axis using an interpolation compensation mechanism to obtain a multidimensional sensing data set; constructing a unified vehicle coordinate system with the centroid of the vehicle to be evaluated as the origin, and mapping the multi-source heterogeneous sensing data to the unified vehicle coordinate system using an external parameter calibration method; and using an online dynamic compensation mechanism to adjust and update the spatial calibration parameters of the vehicle coordinate system according to the attitude changes of the vehicle to be evaluated during operation, so as to ensure the consistency of the multidimensional sensing data set at the spatial level.

[0008] Optionally, forming a multidimensional perception degradation trajectory includes: dividing the environmental disturbances of the vehicle to be evaluated into multiple preset dimensions, increasing the intensity of the disturbances in chronological order, and generating a continuously changing degradation scenario sequence; extracting preset scoring indicators of the vehicle perception evaluation system under each degradation scenario, and using a multidimensional perception degradation function to map the performance changes of the vehicle perception evaluation system to be evaluated into a continuously changing curve, thereby obtaining a multidimensional perception degradation trajectory.

[0009] Optionally, the construction of the credibility reconstruction evaluation model includes: forming a dynamic semantic structure graph based on the interaction relationship between traffic participants and objects in the multidimensional perception degradation trajectory, and decomposing and temporally modeling the target behavior of traffic participants; constructing a multidimensional physical constraint space according to a preset vehicle kinematics model, and performing physical reachability inversion verification on the target trajectory in the multidimensional perception degradation trajectory; when the semantic results conflict with the physical constraints, adopting a credibility reconstruction mechanism to construct a credibility reconstruction evaluation model driven by the semantic-physical consistency constraint mechanism.

[0010] Optionally, the obtained evaluation results include: normalizing the continuous temporal features of the multidimensional perceived degradation trajectory and the discrete structural features in the inversion verification to construct a global evaluation state vector; dynamically adjusting the weight ratios of the degradation trajectory index and the inversion verification index according to the complexity of the adversarial test scenario represented by the global evaluation state vector to construct a multidimensional fusion evaluation function; smoothing the evaluation results of the multidimensional fusion evaluation function through a sliding time window, and introducing an anomaly scoring detection mechanism to correct the evaluation results.

[0011] Optionally, the dynamic generation and optimization of adversarial test scenarios includes: constructing a multi-dimensional perception weakness thermal distribution model based on the performance error of the vehicle perception evaluation system under different target environmental conditions; using the multi-dimensional perception weakness thermal distribution model as a constraint, optimizing the target environmental parameters through a condition generation model to construct adversarial test scenarios; and adaptively adjusting the set of adversarial test scenarios based on the degree of exposure of the defects of the vehicle perception evaluation system under evaluation in the adversarial test scenarios, and eliminating redundant scenarios to achieve dynamic generation and optimization of adversarial test scenarios.

[0012] Optionally, the construction of the closed-loop feedback mechanism includes: performing sequence modeling of the evaluation results of a preset test cycle, and integrating multi-dimensional perception degradation trajectory features and inversion verification features to form cross-cycle historical memory data; constructing a global error function by calculating the deviation between the evaluation results and the actual results, and performing multi-module joint reverse optimization and evaluation parameter update; monitoring the volatility and deviation gradient of the evaluation results to determine whether the perception evaluation system of the vehicle to be evaluated has entered a stable state.

[0013] Optionally, the intelligent driving vehicle perception evaluation method further includes: constructing a global closed-loop system based on multi-dimensional perception degradation trajectory, preset rule constraints, multi-dimensional fusion evaluation function and self-evolving test scenario generation mechanism, and realizing the collaborative evolution of the vehicle perception evaluation system to be evaluated through cross-cycle state unified modeling and error-driven joint optimization.

[0014] Optionally, the construction of a global closed-loop system, through cross-cycle state unified modeling and error-driven joint optimization, enables the collaborative evolution of the vehicle perception evaluation system to be evaluated. This includes: integrating the operational data scattered across various modules of the vehicle perception evaluation system into a global evaluation state vector; introducing a cross-module state alignment mechanism to map the global evaluation state vector of historical versions to the new version system, thereby realizing the transfer of evaluation knowledge; and continuously introducing new environmental disturbance modes to expand the coverage of the degradation trajectory, thus forming a vehicle perception evaluation system with continuous evolution capabilities.

[0015] On the other hand, the present invention provides an intelligent driving vehicle perception evaluation system based on multi-dimensional sensing. The intelligent driving vehicle perception evaluation system includes a control module, the control module including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the intelligent driving vehicle perception evaluation method according to any one of the above.

[0016] Through the above technical solutions, this invention achieves high-precision and consistent expression of multi-dimensional sensing data under different sampling frequencies and spatial coordinate systems by constructing a unified spatiotemporal alignment and fusion processing mechanism for multi-source heterogeneous sensing data, thereby improving the reliability and stability of the basic sensing input data. By establishing a multi-dimensional performance change analysis mechanism based on the influence of environmental disturbances, it achieves continuous characterization and robust boundary identification of the performance change process of the sensing system in complex dynamic traffic scenarios, thereby improving the system's adaptability to complex working conditions and the accuracy of evaluation. By introducing a test environment generation and screening mechanism based on scenario constraints, it achieves automated construction and effective coverage of complex long-tail traffic scenarios, thereby enhancing the verification capability and adaptability of the test system to extreme and rare scenarios.

[0017] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the perception evaluation method for intelligent driving vehicles based on multi-dimensional sensing, as described in this invention. Figure 1 ; Figure 2 This is a schematic diagram of the process for obtaining unified multidimensional sensing data in this invention; Figure 3 This is a schematic diagram of the process for forming a multidimensional sensing degradation trajectory in this invention; Figure 4 This is a schematic diagram of the process for constructing the credibility reconstruction evaluation model in this invention; Figure 5 This is a schematic diagram of the process for obtaining evaluation results in this invention; Figure 6 This is a flowchart illustrating the dynamic generation and optimization of adversarial test scenarios in this invention. Figure 7 This is a schematic diagram of the process for constructing the closed-loop feedback mechanism in this invention; Figure 8 This is a schematic diagram of the process for constructing a global closed-loop system in this invention; Figure 9 This is a flowchart illustrating the perception evaluation method for intelligent driving vehicles based on multi-dimensional sensing, as described in this invention. Figure 2 . Detailed Implementation

[0019] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0020] It should be noted that the acquisition, transmission, storage, use, and processing of data in this application comply with relevant national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0021] Please refer to Figure 1 and Figure 9 This invention provides a perception evaluation method for intelligent driving vehicles based on multi-dimensional sensing. The method may include: Step S100: Based on the multi-dimensional sensors deployed on the vehicle to be evaluated, collect multi-source heterogeneous sensing data, and perform time synchronization and spatial calibration to obtain a unified multi-dimensional sensing data set.

[0022] In a preferred embodiment of the present invention, the multi-dimensional sensing system of the vehicle to be evaluated (e.g., especially high-end models) can be uniformly collected and scheduled, including multi-source sensing devices such as front-view cameras, surround-view cameras, millimeter-wave radar and lidar. Since different sensors have significant differences in sampling frequency, timestamp accuracy and output format, it is necessary to establish a unified time reference synchronization mechanism so that all sensing data are mapped to a unified time axis for processing.

[0023] Please refer to Figure 2In this embodiment of the invention, obtaining unified multidimensional sensing data may include: Step S110: Perform time base unification processing on the multi-source heterogeneous sensing data, and use an interpolation compensation mechanism to align the preset sensor data to the preset time axis to obtain a multi-dimensional sensing data set.

[0024] In a preferred embodiment of the present invention, the master clock synchronization module can be used to trigger and control each sensor to complete data acquisition within the same time window. At the same time, a time interpolation compensation mechanism is introduced for low-frequency sensor data to align it with high-frequency data in the time dimension, thereby avoiding the accumulation of spatial errors caused by time offset. In the data preprocessing stage, the original data is denoised, abnormal frames are removed, and data integrity is verified, ultimately forming a multi-dimensional sensing data set that is strictly consistent in the time dimension.

[0025] Step S120: Construct a unified vehicle coordinate system with the centroid of the vehicle to be evaluated as the origin, and map the multi-source heterogeneous sensor data to the unified vehicle coordinate system through the external parameter calibration method.

[0026] In a preferred embodiment of the present invention, a unified vehicle coordinate system with the centroid of the vehicle to be evaluated as the origin is first constructed, and it is used as the spatial representation benchmark for all sensor data. Subsequently, the rotation matrix and translation vector of each sensor relative to the vehicle coordinate system are obtained through an external parameter calibration method, and the camera image coordinate system, the millimeter-wave radar polar coordinate system, and the lidar three-dimensional point cloud coordinate system are mapped to the vehicle coordinate system. In this process, in order to improve the calibration accuracy, a parameter optimization mechanism based on minimizing reprojection error is introduced, so that the spatial transformation matrix continuously converges iteratively in multiple frames of data.

[0027] In a preferred embodiment of the present invention, the objective function of the parameter optimization mechanism can be represented by the following formula:

[0028] in, This represents the total reprojection error. This represents the actual pixel coordinates of the target point on the image. This represents the corresponding world coordinate system coordinates. and Let these represent the rotation matrix and translation vector (i.e., the extrinsic parameters to be optimized), respectively. These can be solved using the LM iterative algorithm. smallest and .

[0029] Step S130: An online dynamic compensation mechanism is adopted to adjust and update the spatial calibration parameters of the vehicle coordinate system according to the attitude changes of the vehicle under evaluation during operation, so as to ensure the consistency of the multi-dimensional perception data set at the spatial level.

[0030] In a preferred embodiment of the present invention, the online dynamic compensation mechanism is a technical system that maintains system stability and optimal performance by monitoring system status in real time, analyzing deviations, and automatically adjusting parameters. The core of the mechanism is to eliminate environmental interference and system errors through dynamic feedback and adaptive adjustment, so that the system can maintain the best performance in a dynamic environment.

[0031] Step S200: Based on the multi-dimensional perception data set, construct a multi-level environmental disturbance model to generate a multi-dimensional perception degradation scenario, record the output changes of the perception evaluation system of the vehicle to be evaluated under different degradation levels, and form a multi-dimensional perception degradation trajectory.

[0032] Please refer to Figure 3 In this embodiment of the invention, forming a multidimensional sensing degradation trajectory may include: Step S210: Divide the environmental disturbances of the vehicle to be evaluated into multiple preset dimensions, and increase the intensity of the disturbances in chronological order to generate a continuously changing sequence of degradation scenarios.

[0033] In a preferred embodiment of the present invention, the environmental disturbances to the vehicle under evaluation can be divided into three main dimensions: illumination disturbance, weather disturbance, and spatial occlusion disturbance. Regarding illumination disturbance, environmental changes from normal lighting to strong backlight and low nighttime illumination are simulated by adjusting the image brightness curve and contrast parameters. Regarding weather disturbance, fogging, raindrop noise, and snow particle interference models are introduced to simulate the impact of different weather conditions on the sensor. Regarding spatial occlusion disturbance, a random occlusion region generation algorithm simulates vehicle occlusion, pedestrian occlusion, and localized field-of-view loss. These disturbance processes are progressively amplified in chronological order, forming a continuously changing sequence of degradation scenarios, enabling the multi-dimensional sensing system to perform response tests under different environmental pressures.

[0034] Step S220: Extract the preset scoring indicators of the vehicle perception evaluation system under various degradation scenarios, and use a multidimensional perception degradation function to map the performance changes of the vehicle perception evaluation system to a continuous change curve to obtain the multidimensional perception degradation trajectory.

[0035] In a preferred embodiment of the present invention, the target detection results, trajectory prediction results, and semantic recognition results can be uniformly converted into a standardized evaluation vector, including indicators such as detection accuracy, trajectory deviation rate, and semantic consistency score. Simultaneously, a time continuity constraint mechanism is introduced to smooth the output changes between adjacent degradation levels, avoiding evaluation fluctuations caused by random noise. Furthermore, by constructing a multidimensional perceptual degradation function, the system performance changes under different degradation conditions are mapped to a continuously changing curve structure, thereby forming a standardized multidimensional perceptual degradation trajectory, enabling the quantification of the pattern of system performance changes with environmental disturbances.

[0036] In a specific embodiment, in an autonomous driving test scenario, the vehicle sequentially experiences three environmental states on a closed test road: normal weather, light fog, and heavy fog. Under normal weather conditions, the multi-dimensional perception system achieves a 95% accuracy rate in detecting vehicles ahead; this drops to 80% in light fog; and further decreases to 63% in heavy fog. The multi-dimensional perception degradation trajectory model constructed in this step maps performance changes under different environments as a continuous curve and identifies a significant performance inflection point when visibility is below 30 meters, thus accurately characterizing the system's robustness boundary.

[0037] Step S300: Based on the multidimensional perception degradation trajectory, construct a semantic structure graph, and through preset rule constraints, perform inversion verification on the multidimensional perception degradation trajectory to construct a credibility reconstruction evaluation model.

[0038] Please refer to Figure 4 In this embodiment of the invention, constructing a credibility reconstruction evaluation model may include: Step S310: Based on the multidimensional perception degradation trajectory of traffic participants (e.g., vehicles, pedestrians, non-motorized vehicles and static obstacles) and the interaction relationships between them (e.g., spatial distance relationships, speed relationships and interaction influence relationships), a dynamic semantic structure map is formed, and the target behavior of traffic participants is decomposed and temporally modeled.

[0039] In a preferred embodiment of the present invention, vehicles, pedestrians, non-motorized vehicles, and static obstacles can be uniformly abstracted as semantic nodes, and semantic edges can be constructed based on spatial distance relationships, speed relationships, and interaction influence relationships, thereby forming a complete dynamic semantic graph structure. Based on this, a target behavior decomposition mechanism is introduced to break down complex traffic behaviors. For example, "acceleration and lane-changing obstacle avoidance behavior" is decomposed into acceleration sub-behaviors, lateral displacement sub-behaviors, and obstacle avoidance sub-behaviors, and each sub-behavior is independently modeled in time series. Simultaneously, through a semantic consistency propagation mechanism, local semantic errors are detected and diffused within the graph structure, thereby identifying potential semantic conflict areas and improving the structural integrity and interpretability of the semantic expression.

[0040] Step S320: Construct a multi-dimensional physical constraint space (e.g., speed continuity constraint, acceleration upper limit constraint, and steering angle change constraint) based on the preset vehicle kinematics model, and perform physical reachability inversion verification on the target trajectory in the multi-dimensional perception degradation trajectory.

[0041] In a preferred embodiment of the present invention, a time-step mapping analysis can be performed on the target trajectory in the multidimensional perception degradation trajectory, and a physical reachability inversion mechanism can be introduced to perform inverse solution analysis on the target trajectory to determine whether the trajectory can be generated by a real vehicle power system. For example, for trajectories with instantaneous displacement or velocity changes, it is determined whether they exceed the physical limit range by calculating their minimum acceleration requirements.

[0042] In a preferred embodiment of the present invention, the actual lateral acceleration of the vehicle to be evaluated can be expressed by the following formula. :

[0043] in, Indicates vehicle speed. This indicates the turning radius, and a physical limit threshold can be set. (Usually the value is taken as) , (representing gravitational acceleration), if calculated as follows If the trajectory exceeds the physical limit, it is determined that the trajectory is unreachable.

[0044] In addition, to enhance adaptability to complex environments, a probabilistic physical boundary model can be introduced to evaluate confidence intervals for scenarios with ambiguous boundaries, rather than using simple binary judgments, thereby improving the robustness and adaptability of physical constraint judgments.

[0045] Step S330: When semantic results conflict with physical constraints, a credibility reconstruction mechanism is adopted to construct a credibility reconstruction evaluation model driven by semantic-physical consistency constraint mechanism.

[0046] In a preferred embodiment of the invention, the semantic layer output confidence and the degree of physical layer constraint satisfaction can be jointly modeled. A unified confidence scoring system is constructed through a nonlinear mapping function, thereby avoiding the information distortion problem caused by simple linear weighting. When semantic results conflict with physical constraints, they are not directly judged as errors. Instead, a confidence reconstruction mechanism is used to reduce the weight of the conflicting part while retaining the information contribution of the consistent part, thus achieving "soft constraint evaluation". In addition, a historical consistency memory module is introduced to accumulate and learn the semantic-physical consistency performance under similar scenarios, enabling the system to gradually optimize its confidence judgment ability over long-term operation, thereby improving the stability and generalization ability of the evaluation results.

[0047] In a specific embodiment, in a complex traffic test scenario in a city, the system detected a rapid lateral displacement of a vehicle to be evaluated in a short period of time, accompanied by a sudden change in speed. Semantically, this behavior was identified as a "rapid lane change" and a corresponding semantic structure node was generated. However, upon entering the physical constraint verification stage, the system calculated using a vehicle dynamics model that the acceleration requirement corresponding to this trajectory far exceeded the vehicle's physical limits, and the trajectory change did not conform to the laws of continuous motion. Further analysis revealed that this anomaly was caused by a misidentification due to a short-term occlusion by the vehicle in front. Instead of directly removing this result, the system attenuated its semantic weight through a credibility reconstruction mechanism while retaining locally correct identification information, thereby avoiding excessive impact on the overall evaluation system. Ultimately, this output was marked as low credibility in the multi-dimensional fusion evaluation function and entered into the subsequent correction analysis process.

[0048] Step S400: Based on the evaluation results of multidimensional perception degradation trajectory and credibility reconstruction, construct a multidimensional fusion evaluation function containing a dynamic weight adjustment mechanism to obtain the evaluation results.

[0049] Please refer to Figure 5 In this embodiment of the invention, obtaining the evaluation result may include: Step S410: Normalize the continuous temporal features of the multidimensional sensing degradation trajectory and the discrete structural features in the inversion verification to construct a global evaluation state vector.

[0050] In a preferred embodiment of the present invention, since the degradation trajectory index is a continuous time series feature and the semantic-physical consistency index is a structured discrete feature, it is necessary to construct a unified feature representation space across scales. All evaluation indices can be compressed to a unified numerical range using a normalized mapping function, and a multi-scale alignment mechanism can be introduced to enable short-term fluctuation features and long-term performance trend features to be expressed within the same evaluation framework. Simultaneously, all standardized indices are uniformly encoded into a global evaluation state vector to describe the overall performance of the system within a certain testing period. Furthermore, to avoid information coupling interference between different features, a feature decoupling encoding mechanism is introduced, mapping robustness features and consistency features to independent feature subspaces respectively, thereby improving the stability and interpretability of the evaluation results.

[0051] Step S420: Based on the complexity of the adversarial test scenario represented by the global evaluation state vector, dynamically adjust the weight ratios of the degradation trajectory index and the inversion verification index, and construct a multi-dimensional fusion evaluation function.

[0052] In a preferred embodiment of the present invention, the weight distribution between the multidimensional perception degradation trajectory index and the semantic-physical consistency index can be dynamically adjusted according to the complexity characteristics of the current test scenario (e.g., target density, occlusion ratio, velocity change rate, and environmental disturbance intensity). The weight allocation is calculated using the Softmax function.

[0053] In a preferred embodiment of the present invention, the final weight of the evaluation index can be represented by the following formula. :

[0054] in, Indicates the first Evaluation indicators This represents the scene feature vector related to the metric, which includes standardized values ​​such as target density, occlusion ratio, and environmental disturbance intensity. The influence of key features is amplified through an exponential function, so that the weight distribution can be adaptively tilted according to the scene.

[0055] For example, in highly dynamic and complex scenarios, the weight of multi-dimensional perceptual degradation trajectories is increased to emphasize robustness, while in structurally complex but motion-stable scenarios, the weight of semantic-physical consistency is increased. Simultaneously, a weight self-learning mechanism is introduced to continuously optimize the weight parameters through historical evaluation error feedback, enabling the evaluation function to possess long-term adaptive capabilities. Furthermore, to avoid a single indicator dominating the evaluation results, a regularization constraint mechanism is introduced to control the boundary values ​​of the multi-dimensional fusion evaluation function output, thereby ensuring the stability of the scoring results.

[0056] Step S430: Smooth the evaluation results of the multidimensional fusion evaluation function by using a sliding time window, and introduce an anomaly scoring detection mechanism to correct the evaluation results.

[0057] In a preferred embodiment of the invention, a sliding time window can be used to smooth the evaluation results across multiple rounds, thereby mitigating the impact of short-term abnormal fluctuations. Simultaneously, an abnormal scoring detection mechanism is introduced to statistically identify scoring mutation points and, in conjunction with environmental disturbance information, determine whether the fluctuations are reasonable fluctuations caused by extreme scenarios. Abnormal scoring results are not directly removed; instead, a weight decay mechanism is used to correct them, reducing their impact on the overall evaluation. Furthermore, a cross-round consistency verification mechanism is introduced to analyze the trends of scoring results across different testing periods. When abnormal scoring fluctuations are detected, a local calibration mechanism for the evaluation model is triggered to fine-tune and update the weight parameters, thereby improving the stability and robustness of the overall evaluation system.

[0058] In one specific embodiment, during the testing of a high-end autonomous vehicle, the system underwent multiple rounds of testing in both congested urban and highway environments. In the urban congestion scenario, due to the higher target density, the system automatically increased the semantic-physical consistency weight to 0.65, while in the highway scenario, it increased the weight of the multi-dimensional perception degradation trajectory to 0.7. Through this dynamic fusion evaluation mechanism, the system can automatically adjust the evaluation focus according to different scenarios. In the final evaluation results, the comprehensive score for the urban scenario was 0.79, and the score for the highway scenario was 0.74. Compared to traditional fixed-weight evaluation methods, this method can more accurately reflect the true performance differences of the perception system in different scenarios, while avoiding the bias caused by a single indicator dominating the evaluation results.

[0059] Step S500: Based on the evaluation results and the analysis results of the multidimensional perception degradation trajectory, construct a self-evolving test scenario generation mechanism, identify the weak areas of the perception evaluation system of the vehicle to be evaluated and perform reverse scenario modeling, and dynamically generate and optimize adversarial test scenarios.

[0060] Please refer to Figure 6 In this embodiment of the invention, the dynamic generation and optimization of adversarial test scenarios may include: Step S510: Based on the performance error of the vehicle perception evaluation system under different target environmental conditions, construct a multi-dimensional perception weakness thermal distribution model.

[0061] In a preferred embodiment of the invention, environmental variables (e.g., including light intensity, weather disturbances, target density, and motion speed distribution) can be mapped to a unified multidimensional feature space, and corresponding evaluation errors (e.g., including detection errors, trajectory errors, and semantic inconsistency errors) can be embedded in this space to form a high-dimensional error distribution structure. Subsequently, through density clustering analysis and error gradient analysis methods, regions with weak system performance are identified, forming a multidimensional perception weakness region distribution map. Furthermore, a second-order error change rate analysis mechanism is introduced to dynamically characterize the error growth rate, thereby accurately locating the critical region boundary of system performance degradation. Finally, a structured "multidimensional perception weakness thermal distribution model" is output to guide the subsequent test field generation direction.

[0062] Step S520: Using the multidimensional perception weakness thermal distribution model as a constraint, the target environment parameters are optimized in reverse through the condition generation model to construct an adversarial test scenario.

[0063] In a preferred embodiment of the present invention, after identifying weak areas in multidimensional perception, a self-evolving test field generation model based on inversion optimization is constructed. Specifically, using high-risk areas in the error thermal distribution model as constraint inputs, the target environmental parameters of the test field are inversely optimized through a conditional generation model, so that the generated scene can maximize the stimulation of potential defects in the multidimensional perception system.

[0064] In a preferred embodiment of the present invention, the optimization objective of the generative model can be defined by the following formula. :

[0065] in, This represents the loss function to be minimized (the negative sign indicates that the error should be maximized). This indicates the system's perception error (e.g., the rate of decrease in detection rate) in the current generated scene. The weighting coefficients represent the perceived error. This represents the physical plausibility score of the generated scene (e.g., whether the target complies with traffic regulations). The weighting coefficients representing the physical plausibility score can be adjusted using gradient descent to modify environmental parameters (e.g., fog concentration, number of targets). Maximize this to generate high-value adversarial test scenarios.

[0066] Specifically, the target environmental parameters can be decomposed into multiple controllable dimensions, including illumination gradient changes, weather disturbance intensity, target spatial distribution density, and dynamic interaction complexity. The generated scenario is then optimized to gradually approach the system failure boundary. Simultaneously, to ensure the physical rationality of the generated scenario, a physical constraint verification mechanism is introduced to verify the executability of the generated results, avoiding abnormal scenarios that do not conform to real traffic environment patterns. Furthermore, to enhance the diversity of the test field, a random disturbance factor and a diversity regularization mechanism are introduced, enabling the generated test field to maintain its specificity while covering a wide range of scenarios, thereby avoiding overfitting to a single defect pattern.

[0067] Step S530: Based on the degree of exposure of defects in the perception evaluation system of the vehicle under evaluation by the adversarial test scenarios, adaptively adjust the set of adversarial test scenarios and remove redundant scenarios to achieve dynamic generation and optimization of adversarial test scenarios.

[0068] In a preferred embodiment of the invention, the importance of different test environments can be dynamically ranked according to their influence on the multidimensional perception degradation trajectory, and different weights can be assigned to them in the multidimensional fusion evaluation function. For highly sensitive test environments that can significantly expose system defects, their proportion in the overall evaluation is increased, while the weight of low-impact or redundant test environments is reduced, thereby optimizing the allocation of test resources. Simultaneously, a historical feedback-driven evolutionary update mechanism is introduced, feeding test results back to the generation model, enabling continuous iterative optimization of the test environment generation strategy. Furthermore, through redundant scenario elimination and key scenario reinforcement mechanisms, the test environment set maintains a high information density, thereby achieving the continuous self-evolution capability of the testing system.

[0069] In one specific embodiment, during testing of a high-end autonomous driving vehicle, the system discovered through multiple rounds of testing that the false detection rate of the multi-dimensional perception system significantly increased in scenarios involving "low light at night + high density of pedestrians crossing." Based on a self-evolving test field generation mechanism, the system first modeled the error distribution of this type of scenario and identified it as a weak area for multi-dimensional perception. Subsequently, it automatically constructed enhanced test scenarios through a generative model, including adding backlight interference, increasing the dynamic complexity of pedestrians, and introducing irregular motion trajectories. In subsequent tests, this generated scenario successfully triggered multiple false detections by the multi-dimensional perception system, causing the detection accuracy to decrease by approximately 17%. Subsequently, based on test feedback, the system dynamically increased the weight of this type of scenario in the test set, raising its proportion from 10% to 30%, thereby achieving targeted enhancement and continuous optimization of the testing system.

[0070] Step S600: Based on the evaluation results and the target historical data of the multidimensional perception degradation trajectory, construct a closed-loop feedback mechanism to adaptively update the multi-level environmental disturbance model and the credibility reconstruction evaluation model.

[0071] Please refer to Figure 7 In this embodiment of the invention, constructing a closed-loop feedback mechanism may include: Step S610: Perform sequence modeling on the evaluation results of the preset test cycle, and integrate multi-dimensional perception degradation trajectory features and inversion verification features to form cross-cycle historical memory data.

[0072] In a preferred embodiment of the invention, the output values ​​of the multidimensional fusion evaluation function across multiple testing cycles are first uniformly modeled, converting the evaluation results at different time stages into a global evaluation state vector sequence. This state vector not only includes the characteristics of the multidimensional perception degradation trajectory but also integrates the output results of the semantic-physical consistency constraint mechanism and the test field distribution characteristics, thus forming a high-dimensional system state representation. Based on this, a historical memory fusion mechanism is introduced to perform time-weighted processing on the state vectors of different cycles, giving recent data a higher influence weight in the evaluation system. Simultaneously, a sequence modeling method is used to characterize the system performance evolution trend, thereby capturing the changing patterns of the multidimensional perception system during long-term operation.

[0073] Step S620: By calculating the deviation between the evaluation results and the actual results, a global error function is constructed, and multi-module joint reverse optimization and evaluation parameter update are performed.

[0074] In a preferred embodiment of the invention, a global error function can be constructed by calculating the deviation between the evaluation result output by the multidimensional fusion evaluation function and the actual result. This error is then backpropagated to multiple modules, including the multidimensional perception degradation trajectory model, the semantic-physical consistency constraint mechanism, and the self-evolving test field generation mechanism. Unlike traditional single-module optimization methods, a cross-module collaborative optimization strategy is adopted here, enabling mutual constraints and collaborative updates among the modules. For example, when the semantic error increases significantly, not only are the semantic constraint parameters adjusted, but the environmental disturbance intensity parameters in the degradation model are also adjusted simultaneously, thereby achieving system-level consistency optimization. Furthermore, a stability constraint mechanism can be introduced to prevent local optimization from causing overall system oscillations.

[0075] Step S630: Monitor the volatility and deviation gradient of the evaluation results to determine whether the vehicle perception evaluation system under evaluation has entered a stable state.

[0076] In a preferred embodiment of the invention, the system convergence can be determined by monitoring the fluctuation of the output value of the multi-dimensional fusion evaluation function during multiple rounds of testing and combining this with the error gradient changes. When the system score volatility is below a set threshold and the error gradient approaches stability, the system is considered to have entered a convergence state. Simultaneously, to prevent spurious convergence, a structural perturbation verification mechanism is introduced, which verifies the system's stability by introducing a slight perturbation scenario in the test field. Furthermore, a structural stability control module is constructed to dynamically freeze and fine-tune key parameters, thereby ensuring that the system does not experience overfitting or performance degradation during convergence.

[0077] In one specific embodiment, during continuous multi-version testing of a high-end autonomous driving system, the system underwent three rounds of iterative optimization. Through the closed-loop mechanism in this step, the system can automatically perform alignment analysis on the output values ​​of the multi-dimensional fusion evaluation function between different versions and identify the problem of persistently high semantic errors in complex nighttime interaction scenarios. Subsequently, the system automatically triggers a joint optimization mechanism to synchronously adjust the multi-dimensional perception degradation trajectory model and semantic-physical consistency constraint parameters, resulting in an approximately 12% performance improvement for the new version system in subsequent tests. Simultaneously, the system stability is detected through a perturbation verification mechanism. The results show that the score volatility is controlled within ±0.03, indicating that the system has entered a stable convergence state and possesses good cross-cycle consistency.

[0078] Step S700: Based on the multidimensional perception degradation trajectory, preset rule constraints, multidimensional fusion evaluation function and self-evolution test scenario generation mechanism, construct a global closed-loop system, and achieve the collaborative evolution of the perception evaluation system of the vehicle to be evaluated through cross-cycle state unified modeling and error-driven joint optimization.

[0079] Please refer to Figure 8 In this embodiment of the invention, a global closed-loop system is constructed, and the co-evolution of the perception evaluation system of the vehicle to be evaluated is achieved through cross-cycle state unified modeling and error-driven joint optimization. This may include: Step S710: Integrate the operational data scattered across the various modules of the vehicle perception evaluation system into a global evaluation state vector.

[0080] In a preferred embodiment of the present invention, information dispersed across the multidimensional perception degradation trajectory construction module, the semantic-physical consistency constraint mechanism module, the multidimensional fusion evaluation function module, and the self-evolving test field generation mechanism module can be structurally integrated. Specifically, the system operating state of each test cycle is abstracted into a global evaluation state vector. This vector not only includes traditional detection accuracy and robustness indicators but also further integrates high-dimensional information such as degradation trajectory features, semantic consistency scores, physical constraint satisfaction rates, and test field complexity distribution. Based on this, a cross-module state alignment mechanism is introduced, enabling the outputs of different modules to be mapped to a unified state space, thereby forming a comparable system-level expression structure. Simultaneously, through time series modeling methods, the global evaluation state vectors of different test cycles are serialized, allowing the system to capture the dynamic changes in its performance over time, providing basic data support for subsequent closed-loop optimization.

[0081] Step S720: Introduce a cross-module state alignment mechanism to map the global evaluation state vector of the historical version to the new version system, thereby realizing the migration of evaluation knowledge.

[0082] In a preferred embodiment of the invention, a global error function can be constructed by calculating the deviation between the output value of the multidimensional fusion evaluation function and the actual system performance. This error is then backpropagated to multiple functional modules, including a multidimensional perception degradation trajectory model, a semantic-physical consistency constraint mechanism, and a self-evolving test field generation mechanism. During this process, different modules no longer optimize independently but instead form a cross-module collaborative update relationship. For example, when the semantic consistency error increases significantly, the system not only adjusts the semantic parsing parameters but also simultaneously adjusts the environmental disturbance intensity parameters in the degradation trajectory, thereby ensuring the overall consistency of the system. Furthermore, a stability constraint mechanism is introduced to prevent local optimization from causing overall system oscillations, thus ensuring the convergence and robustness of the closed-loop optimization process.

[0083] Step S730: Continuously introduce new environmental disturbance patterns to expand the coverage of degradation trajectories and form a perception evaluation system for the vehicle to be evaluated with continuous evolution capabilities.

[0084] In a preferred embodiment of the invention, the system can be judged to have entered the convergence interval by monitoring the fluctuation range of the multi-dimensional fusion evaluation function output value during multiple rounds of testing and combining the error gradient change trend. When the system score volatility is lower than a set threshold and the error gradient approaches stability, the system is determined to have entered the convergence state. However, to avoid spurious convergence, this step further introduces a structural perturbation verification mechanism. By introducing a slight perturbation scenario in the test field, the stability performance of the system under perturbation conditions is verified. If the system still maintains stable output under perturbation conditions, it is confirmed to be in a true convergence state. In addition, a structural stability control module is introduced to dynamically freeze and fine-tune key parameters, thereby preventing the model from overfitting historical scenarios or local environments and ensuring that the system has long-term stable operation capabilities.

[0085] In a preferred embodiment of the present invention, after the system enters a stable operating phase, the cross-version consistency migration and continuous evolution capability of the multi-dimensional evaluation system can be further realized. Specifically, through a cross-version data alignment mechanism, the global evaluation state vector of the previous version is mapped and aligned with the new version system, making the evaluation system comparable across versions. Simultaneously, a knowledge transfer mechanism is introduced to transfer key experiences from historical test scenarios to the new version evaluation model, thereby reducing the cold start error of the new system. Furthermore, by continuously introducing new test scenarios and new environmental disturbance modes, the coverage of the multi-dimensional perception degradation trajectory is continuously expanded, thereby enhancing the evaluation system's adaptability to unknown scenarios. Ultimately, a multi-dimensional perception evaluation ecosystem with continuous evolution capability is formed, achieving a leap from a static evaluation system to a dynamic self-evolving evaluation system.

[0086] In one specific embodiment, during a six-month continuous test of a high-end autonomous driving system, the system underwent multiple version iterations and updates. Through a cross-version consistency migration mechanism, the system automatically aligns the output values ​​of the multi-dimensional fusion evaluation function across different versions and identifies a decrease in semantic consistency in complex nighttime interaction scenarios in the new version. Subsequently, the system automatically triggers a joint optimization mechanism to synchronously adjust the parameters of the multi-dimensional perception degradation trajectory model and the semantic-physical consistency constraint mechanism, resulting in an approximately 13% performance improvement for the new version in subsequent tests. Simultaneously, a structural perturbation verification mechanism confirms that the system maintains stable output under different test environments, with its score volatility controlled within ±0.03, indicating that the system has entered a true convergence state and possesses good cross-version consistency and continuous evolution capabilities.

[0087] On the other hand, the present invention also provides an intelligent driving vehicle perception evaluation system based on multi-dimensional sensing. The intelligent driving vehicle perception evaluation system includes a control module, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the intelligent driving vehicle perception evaluation method described above.

[0088] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0093] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0094] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0095] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0096] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A perception evaluation method for intelligent driving vehicles based on multi-dimensional sensing, characterized in that, The intelligent driving vehicle perception evaluation method includes: Based on the multi-dimensional sensors deployed on the vehicle to be evaluated, multi-source heterogeneous sensing data is collected, and time synchronization and spatial calibration are performed to obtain a unified multi-dimensional sensing data set. Based on the multidimensional perception data set, a multi-level environmental disturbance model is constructed to generate a multidimensional perception degradation scenario, and the output changes of the perception evaluation system of the vehicle to be evaluated under different degradation levels are recorded to form a multidimensional perception degradation trajectory. Based on the multidimensional perception degradation trajectory, a semantic structure graph is constructed, and the multidimensional perception degradation trajectory is inverted and verified through preset rule constraints to construct a credibility reconstruction evaluation model. Based on the evaluation results of multidimensional perception degradation trajectory and credibility reconstruction, a multidimensional fusion evaluation function containing a dynamic weight adjustment mechanism is constructed to obtain the evaluation results. Based on the evaluation results and the analysis results of the multidimensional perception degradation trajectory, a self-evolving test scenario generation mechanism is constructed to identify the weak areas of the perception evaluation system of the vehicle to be evaluated and perform reverse scenario modeling to dynamically generate and optimize adversarial test scenarios. Based on the evaluation results and the target historical data of the multidimensional perception degradation trajectory, a closed-loop feedback mechanism is constructed to adaptively update the multi-level environmental disturbance model and the credibility reconstruction evaluation model.

2. The intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to claim 1, characterized in that, The unified multidimensional sensing data obtained includes: The time base is unified for multi-source heterogeneous sensing data, and an interpolation compensation mechanism is used to align the preset sensor data to the preset time axis to obtain a multi-dimensional sensing data set. A unified vehicle coordinate system with the centroid of the vehicle to be evaluated as the origin is constructed, and multi-source heterogeneous sensor data are mapped to the unified vehicle coordinate system through the external parameter calibration method. An online dynamic compensation mechanism is adopted to adjust and update the spatial calibration parameters of the vehicle coordinate system according to the attitude changes of the vehicle under evaluation during operation, so as to ensure the consistency of the multi-dimensional perception data set at the spatial level.

3. The intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to claim 1, characterized in that, The formation of the multidimensional sensory degradation trajectory includes: The environmental disturbances to the vehicle to be evaluated are divided into multiple preset dimensions, and the disturbance intensity is increased in chronological order to generate a continuously changing sequence of degradation scenarios. The preset scoring indicators of the vehicle perception evaluation system under various degradation scenarios are extracted, and a multidimensional perception degradation function is used to map the performance changes of the vehicle perception evaluation system under evaluation into a continuous change curve, thus obtaining the multidimensional perception degradation trajectory.

4. The intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to claim 3, characterized in that, The construction of the credibility reconstruction evaluation model includes: Based on the interaction relationship between traffic participants and objects based on the multidimensional perception degradation trajectory, a dynamic semantic structure map is formed, and the target behavior of traffic participants is decomposed and time-series modeled. A multidimensional physical constraint space is constructed based on a pre-defined vehicle kinematics model, and the physical reachability of the target trajectory in the multidimensional perception degradation trajectory is verified by inversion. When semantic results conflict with physical constraints, a credibility reconstruction mechanism is adopted to construct a credibility reconstruction evaluation model driven by semantic-physical consistency constraint mechanism.

5. The intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to claim 1, characterized in that, The evaluation results obtained include: The continuous temporal features of the multidimensional sensing degradation trajectory and the discrete structural features in the inversion verification are normalized to construct a global evaluation state vector; Based on the complexity of the adversarial test scenario represented by the global evaluation state vector, the weight ratios of the degradation trajectory index and the inversion verification index are dynamically adjusted to construct a multi-dimensional fusion evaluation function. The evaluation results of the multidimensional fusion evaluation function are smoothed by using a sliding time window, and an abnormal scoring detection mechanism is introduced to correct the evaluation results.

6. The intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to claim 1, characterized in that, The dynamic generation and optimization of adversarial testing scenarios includes: Based on the performance error of the vehicle perception evaluation system under different target environmental conditions, a multi-dimensional thermal distribution model of perception weaknesses is constructed. Using a multidimensional perception weakness thermal distribution model as a constraint, the target environment parameters are optimized in reverse through a condition generation model to construct an adversarial test scenario; Based on the degree to which the defects of the perception evaluation system of the vehicle under evaluation are exposed in the adversarial test scenarios, the set of adversarial test scenarios is adaptively adjusted and redundant scenarios are eliminated, so as to realize the dynamic generation and optimization of adversarial test scenarios.

7. The intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to claim 1, characterized in that, The construction of the closed-loop feedback mechanism includes: The evaluation results of the preset test cycles are used to perform sequence modeling, and multi-dimensional perception degradation trajectory features and inversion verification features are integrated to form cross-cycle historical memory data; By calculating the deviation between the evaluation results and the actual results, a global error function is constructed, and multi-module joint reverse optimization and evaluation parameter updates are performed. The volatility and deviation gradient of the monitoring and evaluation results are used to determine whether the perception evaluation system of the vehicle under evaluation has entered a stable state.

8. The intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to claim 1, characterized in that, The intelligent driving vehicle perception evaluation method also includes: Based on the multidimensional perception degradation trajectory, preset rule constraints, multidimensional fusion evaluation function and self-evolution test scenario generation mechanism, a global closed-loop system is constructed. Through cross-cycle state unified modeling and error-driven joint optimization, the collaborative evolution of the perception evaluation system of the vehicle to be evaluated is realized.

9. The intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to claim 8, characterized in that, The construction of a global closed-loop system, through cross-cycle state unified modeling and error-driven joint optimization, achieves the collaborative evolution of the perception evaluation system for the vehicle under evaluation, including: The operational data scattered across various modules of the vehicle perception evaluation system to be evaluated are integrated into a global evaluation state vector; A cross-module state alignment mechanism is introduced to map the global evaluation state vector of the historical version to the new version system, thereby realizing the transfer of evaluation knowledge. By continuously introducing new environmental disturbance patterns to expand the coverage of degradation trajectories, a perception evaluation system for vehicles to be evaluated with continuous evolution capabilities can be formed.

10. A perception and evaluation system for intelligent driving vehicles based on multi-dimensional sensing, characterized in that, The intelligent driving vehicle perception evaluation system includes a control module, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the intelligent driving vehicle perception evaluation method based on multi-dimensional sensing according to any one of claims 1-9.

Images