Automatic parking function test method and device, new energy automobile and storage medium

Abstract

The invention relates to an automatic parking function test method and device, a new energy automobile and a storage medium. The method comprises the steps of obtaining real scene data of a vehicle in a preset parking test scene, then injecting the preset parking test scene into an automatic parking system of the vehicle, obtaining vehicle sensing data in an operation process, and performing target entity association based on the vehicle sensing data to obtain an independent target entity list, and carrying out fusion according to the independent target entity list to obtain a fused target entity list, obtaining performance index data of the automatic parking system according to the fused target entity list and the real scene data, and finally, under the condition that the performance index data characterizes that the performance of the automatic parking system does not reach the standard, obtaining the performance index data of the automatic parking system. And step-by-step backtracking and checking are carried out on the fused target entity list, the independent target entity list and the vehicle sensing data to obtain a fault source of the automatic parking system. By adopting the method, the fault positioning accuracy under the condition that the performance does not reach the standard is improved.

Description

Automatic parking function test methods, devices, new energy vehicles and storage media Technical Field

[0001] This application relates to the field of fault location technology, and in particular to an automatic parking function testing method, device, new energy vehicle, computer-readable storage medium, and computer program product. Background Technology

[0002] Most new energy vehicles nowadays are equipped with automatic parking functions. Once the driving is triggered or the automatic parking function is activated, the vehicle can automatically park itself in a parking space based on user control, without human intervention. While different manufacturers may have slight variations in the specific implementation of automatic parking functions, the basic working principle is similar. Generally speaking, automatic parking functions mainly follow three core steps: "perception-planning-execution." Perception refers to the vehicle using its body sensors to detect available parking spaces in the surrounding environment. These sensors can be ultrasonic radar, cameras, millimeter-wave radar, or lidar, or combinations of different types of sensors. Planning involves calculating the parking path. Based on the size and location of the parking space and information about surrounding obstacles (such as other vehicles, pillars, and curbs), the vehicle calculates the optimal parking path. This path typically includes a series of precise instructions, such as moving forward to the starting position, shifting gears, steering, adjusting the angle, and entering the parking space. Execution refers to the system taking over vehicle control after path planning. The control unit uses drive-by-wire technology to issue commands to the steering system (controlling the steering wheel angle), the powertrain system (controlling the throttle and vehicle speed), the braking system (controlling the brakes), and the transmission (controlling forward and reverse gears).

[0003] Throughout the automatic parking process, sensors continuously monitor environmental changes. If a pedestrian, pet, or other obstacle suddenly enters the parking path, the system will automatically apply emergency braking, suspending the parking process until the obstacle is removed, or prompt the driver to take over. However, in some real-world scenarios, when the vehicle is performing automatic parking, the sudden entry of an obstacle into the parking path may cause the vehicle to fail to sense or recognize it quickly enough, resulting in a collision.

[0004] To address the aforementioned issues, existing technologies have proposed various improvement methods, primarily by enhancing vehicle perception capabilities to identify obstacles promptly and accurately, or by optimizing algorithms to precisely identify dynamic obstacles, thereby controlling the vehicle to pause and park, and avoiding collisions.

[0005] However, current testing methods for automatic parking functions suffer from low accuracy in fault location. Summary of the Invention

[0006] Therefore, it is necessary to provide an automatic parking function testing method, device, new energy vehicle, computer-readable storage medium, and computer program product that can improve the accuracy of fault location in response to the above-mentioned technical problems.

[0007] Firstly, this application provides a method for testing an automatic parking function, including:

[0008] Acquire real-world scenario data of the vehicle under preset parking test scenarios;

[0009] The preset parking test scenario is injected into the vehicle's automatic parking system. The vehicle perception data collected by the vehicle's on-board sensors during the operation of the automatic parking system is obtained. Target entities are associated based on each vehicle perception data to obtain an independent target entity list. The vehicle perception data corresponding to each target entity in the independent target entity list is then fused to obtain a fused target entity list.

[0010] Based on the fusion of the target entity list and real-world scenario data, the performance metrics of the automated parking system are obtained;

[0011] When the performance indicators of the automatic parking system fail to meet the standards, the root cause of the failure is found by tracing back the fused target entity list, the independent target entity list, and the perception data of each vehicle step by step.

[0012] In conjunction with the first aspect, in one embodiment, the root cause of the automatic parking system's fault is obtained by tracing back the fused target entity list, the independent target entity list, and the perception data of each vehicle level, including:

[0013] The data quality of the first data quality of the fused target entity list, the second data quality of the independent target entity list, and the third data quality of the perception data of each vehicle are obtained.

[0014] Based on the first data quality, the second data quality, and the third data quality, the cause of the automatic parking system failure is obtained.

[0015] Root cause analysis was performed on the fault to identify the root cause of the automatic parking system failure.

[0016] In conjunction with the first aspect, in one embodiment, the cause of failure of the automatic parking system is obtained based on the first data quality, the second data quality, and the third data quality, including:

[0017] If the initial data quality meets the standards, it is determined that the decision-making or control module of the automatic parking system has malfunctioned.

[0018] If the quality of the first data is substandard but the quality of the second data is satisfactory, it is determined that the fusion algorithm of the automatic parking system is faulty.

[0019] If the second data quality is substandard, the cause of the automatic parking system failure is determined based on the second and third data quality.

[0020] In conjunction with the first aspect, in an exemplary embodiment, the cause of failure in the automatic parking system is determined based on the second and third data quality, including:

[0021] From the list of independent target entities, identify the perception data of faulty vehicles that do not meet the second data quality standard;

[0022] Based on the target vehicle sensor corresponding to the faulty vehicle perception data, obtain the target vehicle perception data of the target vehicle sensor from the perception data of each vehicle.

[0023] If the third data quality of the target vehicle perception data meets the standard, it is determined that the target perception algorithm of the target vehicle sensor is faulty.

[0024] If the quality of the third data of the target vehicle's perception data is substandard, it is determined that the input conditions are insufficient or the hardware of the target vehicle's onboard sensor is faulty.

[0025] In conjunction with the first aspect, in one embodiment, root cause analysis is performed on the cause of the failure to obtain the root cause of the automatic parking system failure, including:

[0026] If the cause of the failure is a fault in the target perception algorithm, the first preliminary conclusion is that the target perception model has insufficient generalization ability in the preset parking test scenario, and the model training data of the target perception model is checked.

[0027] When the inspection results indicate that the training data for the target perception model lacks data under the preset parking test scenario, the root cause of the automatic parking system failure is determined to be the defect in the training data.

[0028] Given that the training data for the target perception model does not lack data from the preset parking test scenario, the second preliminary conclusion is that the model architecture or algorithm for generating the target perception model has limitations, and the intermediate layer output of the target perception model is analyzed.

[0029] Based on the analysis results characterizing the target perception model's attention error features, the root cause of the automatic parking system's failure was determined to be a defect in the model's attention mechanism.

[0030] Given that the analysis results indicate that the target perception model did not focus on error features, a third preliminary conclusion is generated that the vehicle perception data preprocessing stage introduces errors, and the vehicle perception data and the preprocessed vehicle perception data are compared.

[0031] When the comparison results indicate that the vehicle perception enhancement parameters are improperly configured, the root cause of the automatic parking system failure is determined to be an incorrect configuration of the data preprocessing pipeline.

[0032] In conjunction with the first aspect, in one embodiment, performance index data of the automatic parking system is obtained based on the fused target entity list and real-world scene data, including:

[0033] Obtain the correlation between each fusion target entity in the fusion target entity list and each physical entity in the real scene data;

[0034] Based on the correlation between each fusion target entity and each physical entity, real associated target pairs, first virtual associated target pairs, and second virtual associated target pairs are constructed; real associated target pairs include fusion target entities and physical entities, first virtual associated target pairs include virtual fusion target entities and physical entities, and second virtual associated target pairs include fusion target entities and virtual physical entities.

[0035] Based on the real associated target pairs, the first virtual associated target pairs, and the second virtual associated target pairs, the performance index data of the automatic parking system is obtained.

[0036] In conjunction with the first aspect, in an exemplary embodiment, the method further includes:

[0037] For any given performance indicator data, if the performance indicator data is less than a preset performance indicator data threshold, the automatic parking system is determined to be substandard.

[0038] Alternatively, based on the pre-set weights for each performance indicator, the data of each performance indicator is weighted and summed to obtain the overall performance indicator data of the automatic parking system. If the overall performance indicator data is less than the preset overall performance indicator data threshold, the performance of the automatic parking system is determined to be substandard.

[0039] In conjunction with the first aspect, in one embodiment, performance metrics data include recall, precision, F1 score, and classification accuracy; based on real-world associated target pairs, a first virtual associated target pair, and a second virtual associated target pair, performance metrics data for the automated parking system are obtained, including:

[0040] Based on real associated target pairs and the first virtual associated target pair, the recall rate of the automatic parking system is obtained;

[0041] The accuracy of the automatic parking system is obtained based on the real associated target pairs and the second virtual associated target pairs.

[0042] The F1 score of the automatic parking system is obtained based on recall rate and precision rate;

[0043] The classification accuracy of the automatic parking system is obtained based on the classification information of the real associated target pairs and the classification information of the physical entities in the real associated target pairs.

[0044] In conjunction with the first aspect, in one embodiment, vehicle perception data corresponding to each target entity in the independent target entity list are fused to obtain a fused target entity list, including:

[0045] The vehicle perception data corresponding to each target entity in the list of independent target entities is input into a pre-set Kalman filter for position fusion to obtain the fused position of each target entity;

[0046] Based on the confidence level set for the vehicle-mounted sensor, the fusion confidence level of each target entity is obtained;

[0047] By combining the vehicle perception data, fused location, and fused confidence of each target entity, a list of fused target entities is obtained.

[0048] Secondly, this application also provides an automatic parking function testing device, comprising:

[0049] The data acquisition module is used to acquire real-world scenario data of the vehicle under preset parking test scenarios;

[0050] The parking test module is used to inject a preset parking test scenario into the vehicle's automatic parking system, obtain vehicle perception data collected by the vehicle's on-board sensors during the operation of the automatic parking system, associate target entities based on each vehicle perception data to obtain an independent target entity list, and fuse the vehicle perception data corresponding to each target entity in the independent target entity list to obtain a fused target entity list.

[0051] The indicator generation module is used to obtain performance indicator data of the automatic parking system based on the fusion target entity list and real scene data;

[0052] The backtracking and troubleshooting module is used to backtrack and troubleshoot the fused target entity list, the independent target entity list, and the perception data of each vehicle step by step when the performance index data indicates that the automatic parking system's performance is not up to standard, in order to find the root cause of the automatic parking system's failure.

[0053] Thirdly, this application also provides a new energy vehicle, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0054] Acquire real-world scenario data of the vehicle under preset parking test scenarios;

[0055] The preset parking test scenario is injected into the vehicle's automatic parking system. The vehicle perception data collected by the vehicle's on-board sensors during the operation of the automatic parking system is obtained. Target entities are associated based on each vehicle perception data to obtain an independent target entity list. The vehicle perception data corresponding to each target entity in the independent target entity list is then fused to obtain a fused target entity list.

[0056] Based on the fusion of the target entity list and real-world scenario data, the performance metrics of the automated parking system are obtained;

[0057] When the performance indicators of the automatic parking system fail to meet the standards, the root cause of the failure is found by tracing back the fused target entity list, the independent target entity list, and the perception data of each vehicle step by step.

[0058] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0059] Acquire real-world scenario data of the vehicle under preset parking test scenarios;

[0060] The preset parking test scenario is injected into the vehicle's automatic parking system. The vehicle perception data collected by the vehicle's on-board sensors during the operation of the automatic parking system is obtained. Target entities are associated based on each vehicle perception data to obtain an independent target entity list. The vehicle perception data corresponding to each target entity in the independent target entity list is then fused to obtain a fused target entity list.

[0061] Based on the fusion of the target entity list and real-world scenario data, the performance metrics of the automated parking system are obtained;

[0062] When the performance indicators of the automatic parking system fail to meet the standards, the root cause of the failure is found by tracing back the fused target entity list, the independent target entity list, and the perception data of each vehicle step by step.

[0063] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0064] Acquire real-world scenario data of the vehicle under preset parking test scenarios;

[0065] The preset parking test scenario is injected into the vehicle's automatic parking system. The vehicle perception data collected by the vehicle's on-board sensors during the operation of the automatic parking system is obtained. Target entities are associated based on each vehicle perception data to obtain an independent target entity list. The vehicle perception data corresponding to each target entity in the independent target entity list is then fused to obtain a fused target entity list.

[0066] Based on the fusion of the target entity list and real-world scenario data, the performance metrics of the automated parking system are obtained;

[0067] When the performance indicators of the automatic parking system fail to meet the standards, the root cause of the failure is found by tracing back the fused target entity list, the independent target entity list, and the perception data of each vehicle step by step.

[0068] The aforementioned automatic parking function testing method, device, new energy vehicle, computer-readable storage medium, and computer program product acquire real-world scenario data of the vehicle under a preset parking test scenario. This preset parking test scenario is then injected into the vehicle's automatic parking system. During operation, the system acquires vehicle perception data collected by the vehicle's onboard sensors. Based on this vehicle perception data, target entities are associated to obtain an independent target entity list. The vehicle perception data corresponding to each target entity in the independent target entity list is then fused to obtain a fused target entity list. Based on the fused target entity list and the real-world scenario data, performance index data of the automatic parking system is obtained. Finally, if the performance index data indicates that the automatic parking system's performance is substandard, the fused target entity list, the independent target entity list, and each vehicle perception data are traced back level by level to identify the root cause of the automatic parking system's fault. By utilizing the preset scenario injection and multi-layer data backtracking mechanism, performance judgment and fault location are combined, thereby improving the accuracy of fault location when performance is substandard. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 is an application environment diagram of an automatic parking function test method in one embodiment;

[0071] Figure 2 is a flowchart illustrating an automatic parking function testing method in one embodiment;

[0072] Figure 3 is a system architecture diagram of the automatic parking function of a vehicle in one embodiment;

[0073] Figure 4 is a schematic diagram of the data fusion process in another embodiment;

[0074] Figure 5 is a root cause analysis flowchart for the problem "the camera perception algorithm misidentifies behavior as a vehicle" in one embodiment;

[0075] Figure 6 is a structural block diagram of an automatic parking function test device in one embodiment;

[0076] Figure 7 is an internal structure diagram of a new energy vehicle in one embodiment. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0078] The automatic parking testing method provided in this application embodiment can be applied to the application environment shown in Figure 1. The vehicle communicates with the cloud 104 via a network. The data storage system can store the data that the cloud 104 needs to process. The data storage system can be integrated on the cloud 104, or it can be located on the cloud or other network servers. After vehicle 102 arrives at the designated location, the user initiates an automatic parking request to cloud 104 via a mobile terminal such as a smartphone. Cloud 104 then issues an automatic parking command to vehicle 102. In response, vehicle 102 acquires real-world data from a preset parking test scenario and injects this scenario into the vehicle's automatic parking system. It then acquires vehicle perception data collected by the vehicle's onboard sensors during system operation and associates these perception data with target entities to obtain an independent target entity list. Finally, it fuses the vehicle perception data corresponding to each target entity in the independent target entity list to obtain a fused target entity list. Based on the fused target entity list and the real-world data, it obtains performance metrics for the automatic parking system. If the performance metrics indicate that the automatic parking system's performance is substandard, it performs a step-by-step investigation of the fused target entity list, the independent target entity list, and the vehicle perception data to identify the root cause of the automatic parking system's failure. Cloud 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0079] In an exemplary embodiment, as shown in FIG2, an automatic parking function testing method is provided. Taking the application of this method to vehicle 102 in FIG1 as an example, the method includes the following steps S201 to S204. Wherein:

[0080] Step S201: Obtain real-world scenario data of the vehicle in a preset parking test scenario.

[0081] Among them, parking test scenarios can be understood as the process of parking a vehicle in different environments, and real-world scenario data can be understood as the real physical entities that appear outside the vehicle during the parking process.

[0082] For example, after vehicle 102 reaches the designated location and is no longer needed, the user closes the vehicle and gets out, and initiates an automatic parking request to cloud 104 via a mobile terminal such as a mobile phone. Cloud 104 sends an automatic parking command to vehicle 102, and vehicle 102 responds to the automatic parking command by performing an automatic parking operation. The vehicle 102 also records real scene data under the preset parking test scenario through the on-board sensors associated with vehicle 102, including image data collected by cameras and radar data collected by lidar.

[0083] Step S202: Inject the preset parking test scenario into the vehicle's automatic parking system, obtain the vehicle perception data collected by the vehicle's on-board sensors during the operation of the automatic parking system, associate target entities based on each vehicle perception data to obtain an independent target entity list, and fuse the vehicle perception data corresponding to each target entity in the independent target entity list to obtain a fused target entity list.

[0084] Among them, vehicle-mounted sensors can be understood as data sensors associated with a vehicle, including cameras, radar, lidar, etc. Target entities can be understood as physical objects that are identified, classified, and tracked by sensors and algorithms, serving as the basic unit for environmental perception.

[0085] Optionally, a preset test scenario is applied to the automatic parking system of vehicle 102 through hardware-in-the-loop simulation testing or real vehicle testing, and the original vehicle perception data collected by the vehicle's on-board sensors during the operation of the automatic parking system is acquired. The original vehicle perception data is then synchronized in time and space. Specifically, the vehicle perception data is collected at the same time point, and the collected vehicle perception data is unified to a spatial coordinate system with the vehicle as the origin (i.e., X-forward direction, Y-left direction, Z-upward direction) to obtain the vehicle perception data of each vehicle. Then, it is further determined which targets in the vehicle perception data from different on-board sensors are actually the same target entity, and the target entity and the corresponding vehicle perception data are integrated to obtain an independent target entity list. Finally, according to the vehicle perception data corresponding to each target entity in the independent target entity list, position fusion is performed through a Kalman filter to obtain the fused position of each target entity, and the confidence of each on-board sensor is fused to obtain the fused confidence of each target entity. The vehicle perception data, fused position, and fused confidence of each target entity are combined to obtain a fused target entity list.

[0086] Step S203: Based on the fusion target entity list and real-world scenario data, obtain the performance index data of the automatic parking system.

[0087] Among them, performance index data can be understood as a unified quantitative evaluation information between simulation data in the simulation test scenario and real data in the preset test scenario.

[0088] For example, vehicle 102 associates each fusion target entity in the fusion target entity list with the real physical entity in the real scene data, that is, performs target pairing, and obtains positive examples of successful pairing, first negative examples of no corresponding real physical entity in the case of failed pairing, and second negative examples of no corresponding fusion target entity. The performance index calculation is performed according to the positive examples, the first negative examples and the second negative examples to obtain the performance index data of the automatic parking system.

[0089] Step S204: If the performance index data indicates that the automatic parking system's performance is substandard, the fused target entity list, the independent target entity list, and the perception data of each vehicle are traced back level by level to find the root cause of the automatic parking system's failure.

[0090] The root cause of the malfunction can be understood as the specific reason for the parking failure.

[0091] Optionally, if any performance index data is less than a preset index data threshold, the performance of the automatic parking system is determined to be substandard. Alternatively, based on the index weights corresponding to each performance index data, the performance index data is weighted and summed to obtain the corresponding overall performance index data. If the overall performance index data is less than a preset overall performance index data threshold, the performance of the automatic parking system is determined to be substandard. Subsequently, the fused target entity list, the independent target entity list, and the perception data of each vehicle are backtracked and investigated step by step. The cause of the failure that causes the automatic parking system to be substandard is determined by the elimination method. Finally, the root cause analysis is performed on the cause of the failure to obtain the root cause of the automatic parking system failure.

[0092] In the aforementioned automatic parking function testing method, real-world scenario data of the vehicle under a preset parking test scenario is acquired. This preset parking test scenario is then injected into the vehicle's automatic parking system. Vehicle perception data collected by the vehicle's onboard sensors during operation is obtained, and target entities are associated based on this data to obtain an independent target entity list. The vehicle perception data corresponding to each target entity in the independent target entity list is then fused to obtain a fused target entity list. Based on the fused target entity list and the real-world scenario data, performance index data of the automatic parking system is obtained. Finally, if the performance index data indicates that the automatic parking system's performance is substandard, the fused target entity list, the independent target entity list, and each vehicle perception data are traced back level by level to identify the root cause of the automatic parking system's fault. By utilizing the preset scenario injection and multi-layer data backtracking mechanism, performance judgment and fault location are combined, thereby improving the accuracy of fault location when performance is substandard.

[0093] In one embodiment, the root cause of the automatic parking system failure is obtained by tracing back the fused target entity list, the independent target entity list, and the perception data of each vehicle level. This includes: obtaining the first data quality of the fused target entity list, the second data quality of the independent target entity list, and the third data quality of the perception data of each vehicle; obtaining the cause of the automatic parking system failure based on the first data quality, the second data quality, and the third data quality; and performing root cause analysis on the cause of the failure to obtain the root cause of the automatic parking system failure.

[0094] The first data quality can be understood as the correctness and completeness of each fused target entity in the fused target entity list; the second data quality can be understood as the correctness of the vehicle perception data corresponding to the independent target entities in the independent target entity list; and the third data quality can be understood as the quality of the vehicle perception data itself, such as image quality and other indicators.

[0095] For example, vehicle 102 performs a layer-by-layer backtracking analysis in the order of fusion layer → target layer → original layer. First, it obtains the first data quality of the fusion target entity list, the second data quality of the independent target entity list, and the third data quality of the perception data of each vehicle. Then, it performs fault troubleshooting of the fusion layer based on the first data quality, fault troubleshooting of the target layer based on the second data quality, and fault troubleshooting of the original layer based on the third data quality, thereby obtaining the fault cause of the automatic parking system. Subsequently, the fault cause is subjected to root cause analysis to obtain the root cause of the automatic parking system fault.

[0096] Based on the aforementioned implementation method, the fault is eliminated layer by layer by tracing back step by step, thereby ensuring the accuracy of the fault cause determination. Then, further root cause analysis is performed on the fault to obtain the root cause of the automatic parking system fault. This allows relevant personnel to make corresponding improvement measures based on the root cause of the fault, thereby improving the parking accuracy of the vehicle and enhancing the user experience.

[0097] In one embodiment, the cause of failure of the automatic parking system is obtained based on the first data quality, the second data quality, and the third data quality, including: if the first data quality meets the standard, determining that the decision or control module of the automatic parking system is faulty; if the first data quality is substandard but the second data quality meets the standard, determining that the fusion algorithm of the automatic parking system is faulty; if the second data quality is substandard, determining the cause of failure of the automatic parking system based on the second data quality and the third data quality.

[0098] Optionally, if all fusion target entities are correct and complete, i.e., the first data quality meets the standard, vehicle 102 determines that the decision or control module of the automatic parking system has malfunctioned. If any fusion target entity is incorrect or incomplete, i.e., the first data quality does not meet the standard, and the second data quality of the independent target entity corresponding to the fusion target entity meets the standard, vehicle 102 determines that the fusion algorithm of the automatic parking system has malfunctioned. If the second data quality of any independent target entity does not meet the standard, the cause of the automatic parking system malfunction is determined based on the second data quality of the independent target entity and the third data quality corresponding to the vehicle perception data of the independent target entity in the original layer.

[0099] According to the above implementation method, the refined process of backtracking and cross-validation ensures the accuracy of fault location, greatly shortens the fault investigation cycle, and ensures the timeliness of vehicle maintenance, thereby improving the user experience and enhancing user stickiness.

[0100] In an exemplary embodiment, determining the cause of a malfunction in the automatic parking system based on second and third data quality includes: identifying faulty vehicle perception data with substandard second data quality from a list of independent target entities; obtaining target vehicle perception data of the target vehicle sensor from each vehicle perception data according to the target vehicle sensor corresponding to the faulty vehicle perception data; determining that the target perception algorithm of the target vehicle sensor is faulty if the third data quality of the target vehicle perception data is satisfactory; and determining that the input conditions are insufficient or the hardware of the target vehicle sensor is faulty if the third data quality of the target vehicle perception data is substandard.

[0101] For example, vehicle 102 determines faulty vehicle perception data with substandard second data quality from a list of independent target entities, and then obtains the target vehicle perception data of the target vehicle sensor from each vehicle perception data based on the target vehicle sensor corresponding to the faulty vehicle perception data. If the third data quality of the target vehicle perception data meets the standard, vehicle 102 determines that the target perception algorithm of the target vehicle sensor is faulty. If the third data quality of the target vehicle perception data is substandard, vehicle 102 determines that the input conditions or the hardware of the target vehicle sensor is faulty.

[0102] Based on the aforementioned implementation methods, through the meticulous process of backtracking and cross-validation at each level, fault location achieves the isolation of problems clearly to hardware sensors or software algorithms, the accuracy of locating specific processing steps within the sensor (such as classification, detection, and signal filtering), and the diagnostic conclusions derived from synchronously recorded multi-layered data, which are verifiable and traceable, thereby improving the reliability of fault location.

[0103] In one embodiment, root cause analysis is performed on the cause of the failure to obtain the root cause of the automatic parking system failure, including: if the failure is due to a fault in the target perception algorithm, a first preliminary conclusion is generated that the target perception model has insufficient generalization ability in a preset parking test scenario, and the model training data of the target perception model is checked; if the check results indicate that the model training data of the target perception model lacks data in the preset parking test scenario, the root cause of the automatic parking system failure is determined to be a defect in the training data; if the check results indicate that the model training data of the target perception model does not lack data in the preset parking test scenario, a second preliminary conclusion is generated that the model architecture or algorithm of the target perception model has limitations, and the intermediate layer output of the target perception model is analyzed; if the analysis results indicate that the target perception model focuses on erroneous features, the root cause of the automatic parking system failure is determined to be a defect in the model attention mechanism; if the analysis results indicate that the target perception model does not focus on erroneous features, a third preliminary conclusion is generated that the vehicle perception data preprocessing stage introduces errors, and the vehicle perception data and the preprocessed vehicle perception data are compared; if the comparison results indicate that the vehicle perception enhancement parameters are improperly configured, the root cause of the automatic parking system failure is determined to be an error in the data preprocessing pipeline configuration.

[0104] Among them, the target perception model can be understood as a reinforcement learning model deployed on the target vehicle perception system to perform target perception.

[0105] Optionally, if the fault is caused by a malfunction in the target perception algorithm of the onboard vehicle, vehicle 102 first makes a preliminary conclusion that the generalization ability of the target perception model in the preset parking test scenario is insufficient. Then, guided by the first preliminary conclusion, it checks the model training data of the target perception model. If the check results indicate that the model training data of the target perception model lacks data in the preset parking scenario, the root cause of the automatic parking system fault is determined to be a defect in the training data. If the check results indicate that the model training of the target perception model does not lack data in the preset parking scenario, vehicle 102 makes a second preliminary conclusion that the model architecture or algorithm of the target perception model itself has limitations, and uses the second preliminary conclusion as the basis for its determination. Guided by the preliminary conclusions, the intermediate layer output of the target perception model is analyzed. If the analysis results indicate that the target perception model focuses on incorrect features, the root cause of the automatic parking system's failure is determined to be a defect in the model's attention mechanism. If the analysis results indicate that the target perception does not focus on incorrect features, Vehicle 102 makes a third preliminary conclusion that errors are introduced in the vehicle perception data preprocessing stage. Guided by this third preliminary conclusion, the vehicle perception data and the preprocessed vehicle perception data are compared. If the comparison results indicate that the vehicle perception enhancement parameters are improperly configured, the data preprocessing pipeline configuration error is taken as the root cause of the automatic parking system's failure. Based on the aforementioned diagnostic results, a final diagnostic result is generated and corresponding rectification measures are formulated.

[0106] According to the above implementation method, by drawing conclusions and performing root cause analysis in a step-by-step manner, the accuracy of root cause analysis is improved. Furthermore, by generating diagnostic reports and formulating corresponding rectification measures, it is convenient for development and maintenance personnel to grasp the vehicle fault status and root cause of the fault, thereby judging whether the rectification measures are in place and successfully eliminating the corresponding fault, thus enabling the normal execution of related functions.

[0107] In one embodiment, performance index data of the automatic parking system is obtained based on the fusion target entity list and real scene data, including: obtaining the correlation degree between each fusion target entity in the fusion target entity list and each physical entity in the real scene data; constructing real associated target pairs, a first virtual associated target pair, and a second virtual associated target pair based on the correlation degree between each fusion target entity and each physical entity; the real associated target pair includes a fusion target entity and a physical entity, the first virtual associated target pair includes a virtual fusion target entity and a physical entity, and the second virtual associated target pair includes a fusion target entity and a virtual physical entity; and obtaining performance index data of the automatic parking system based on the real associated target pairs, the first virtual associated target pairs, and the second virtual associated target pairs.

[0108] Among them, the degree of correlation can be understood as the quantitative cost information that the target entity and the real physical entity may be the same entity. The real related target pair can be understood as the real combination of the target entity and the real physical entity that corresponds to the real physical entity. The first virtual related target pair can be understood as the virtual combination that has a real physical entity but does not correspond to the target entity. Similarly, the second virtual related target pair can be understood as the virtual combination that does not correspond to the real physical entity.

[0109] For example, vehicle 102 uses the Hungarian algorithm to obtain the association cost between each fusion target entity in the fusion target entity list and each physical entity in the real scene data. Simultaneously, an association threshold is set; only target pairs with association costs lower than the threshold are successfully associated, meaning the fusion target entity and the physical entity on the ground are successfully paired, forming a real association pair. If the association fails, it may be due to a missed detection—the ground target did not find a matching fusion target, meaning the system did not detect this real object, forming a first virtual association target pair; or it may be due to a false detection—the fusion target did not find a matching ground target, meaning the system fabricated a non-existent object, forming a second virtual association target pair. Finally, based on the real association target pairs, the first virtual association target pairs, and the third virtual association target pairs, the performance index data of the automatic parking system is obtained.

[0110] Based on the aforementioned implementation method, by associating targets to form positive examples, negative examples 1 and negative examples 2, it is convenient to perform quantitative processing of system performance indicators based on the constructed positive and negative examples, thereby accelerating the calculation speed of performance indicator data.

[0111] In an exemplary embodiment, the method further includes: for any performance index data, if the performance index data is less than a preset performance index data threshold, determining that the performance of the automatic parking system is substandard; or, based on the index weights preset for each performance index data, performing a weighted summation of each performance index data to obtain the overall performance index data of the automatic parking system, and if the overall performance index data is less than a preset overall performance index data threshold, determining that the performance of the automatic parking system is substandard.

[0112] Optionally, for any performance index data, if the performance index data is less than a preset performance index data threshold, vehicle 102 determines that the performance of the automatic parking system is substandard; or, obtain the index weights preset for each performance index data, and perform a weighted summation of each performance index data based on each index weight to obtain the overall performance index data of the automatic parking system. If the overall performance index data is less than a preset overall performance index data threshold, vehicle 102 determines that the performance of the automatic parking system is substandard.

[0113] According to the above implementation method, a "one-vote veto" system can be used to execute performance adjudication, or a comprehensive scoring system can be used for performance adjudication. This allows for the selection of adjudication methods based on actual circumstances, thereby ensuring the accuracy of performance adjudication.

[0114] In one embodiment, performance metrics data include recall, precision, F1 score, and classification accuracy. The performance metrics data of the automatic parking system are obtained based on real-object pairs, a first virtual-object pair, and a second virtual-object pair, including: obtaining the recall of the automatic parking system based on the real-object pairs and the first virtual-object pair; obtaining the precision of the automatic parking system based on the real-object pairs and the second virtual-object pair; obtaining the F1 score of the automatic parking system based on the recall and precision; and obtaining the classification accuracy of the automatic parking system based on the classification information of the real-object pairs and the classification information of the physical entities in the real-object pairs.

[0115] Among them, recall rate can be understood as a quantitative information that measures the ability of an automatic parking system to find real objects; precision rate can be understood as a quantitative information that measures whether the detection results of an automatic parking system are accurate and reliable; F1 score can be understood as the harmonic mean of recall rate and precision rate, a comprehensive indicator that measures detection performance.

[0116] For example, vehicle 102 calculates the recall rate of the automatic parking system based on real associated target pairs and a first virtual associated target pair, calculates the precision of the automatic parking system based on real associated target pairs and a second virtual associated target pair, calculates the harmonic mean of recall rate and precision rate to obtain the F1 value of the automatic parking system, and finally calculates the classification accuracy of the automatic parking system based on the classification information of real associated target pairs and the classification information of physical entities in real associated target pairs.

[0117] Based on the aforementioned implementation methods, by calculating recall rate, precision rate, F1 score, and classification accuracy, a comprehensive system measurement of the automatic parking system is achieved, ensuring the comprehensiveness and accuracy of the system measurement.

[0118] In one embodiment, a fused target entity list is obtained by fusing vehicle perception data corresponding to each target entity in the independent target entity list. This includes: inputting the vehicle perception data corresponding to each target entity in the independent target entity list into a pre-set Kalman filter for position fusion to obtain the fused position of each target entity; obtaining the fusion confidence of each target entity based on the confidence level set for the vehicle sensor; and combining the vehicle perception data, fused position, and fusion confidence of each target entity to obtain the fused target entity list.

[0119] The Kalman filter can be understood as an algorithm used to estimate the state of a dynamic system.

[0120] Optionally, the vehicle 102 inputs the vehicle perception data corresponding to each target entity in the independent target entity list into a pre-set Kalman filter for position fusion to obtain the fused position of each target entity, and obtains the fusion confidence of each target entity based on the confidence level set for the vehicle-mounted sensor. Finally, the vehicle perception data, fused position and fusion confidence of each target entity are combined to obtain the fused target entity list.

[0121] According to the above implementation method, position fusion is performed by Kalman filter, which improves the accuracy of fused position. By calculating the fusion confidence and using it as a component of the fused target entity list, the information richness of the fused target entity list is increased, thus ensuring the credibility of the fused target entity list.

[0122] In an exemplary embodiment, as shown in Figure 3, a system architecture for a vehicle automation function is provided, wherein the system architecture for the vehicle automatic parking function consists of a vehicle-side system, a site-side system, and a cloud system:

[0123] Vehicle-side system architecture:

[0124] The vehicle terminal system consists of a perception layer, a decision-making layer, and an execution layer.

[0125] The perception layer includes onboard sensors and a positioning module. Ultrasonic radar (for close-range parking), surround-view cameras (for identifying parking lines and obstacles), millimeter-wave radar (for speed and distance measurement with strong anti-interference capabilities), and lidar (used in some high-end solutions for high-precision 3D environmental modeling) are all onboard sensors that provide real-time data on the vehicle's surroundings. The positioning module includes high-precision GPS + IMU (Inertial Measurement Unit) and wheel speedometers, providing initial vehicle position information. However, in scenarios where GPS is prone to failure, such as underground parking lots, on-site systems are needed to assist in positioning.

[0126] The decision-making layer consists of an onboard computing platform and a high-precision map. The onboard computing platform is a high-performance computing unit that runs complex autonomous driving algorithms. It combines information transmitted from the vehicle and parking lot systems to perform global path planning (the approximate route from the current location to the target parking space) and local path planning (real-time obstacle avoidance, following other vehicles, and other fine-grained operations). In addition, the vehicle carries a high-precision map of the parking lot, containing detailed information such as the precise location of lanes, slope, curvature, parking space locations, no-parking zones, and speed bumps, which forms the basis for vehicle positioning and parking planning.

[0127] The execution layer precisely executes the instructions issued by the computing platform through a drive-by-wire system consisting of drive-by-wire, steering-by-wire, and brake-by-wire, controlling the vehicle's throttle, brakes, steering wheel, and gears.

[0128] Field-end system architecture:

[0129] The field infrastructure consists of a sensing layer and a computing and communication layer.

[0130] The perception layer includes a camera network and LiDAR. The camera network is distributed throughout the parking lot ceiling and key nodes, providing a global, blind-spot-free monitoring perspective, and can detect vehicles, pedestrians, obstacles, and vacant parking spaces. In addition, in some solutions, LiDAR is also deployed at the site to create a more accurate real-time dynamic environment.

[0131] The computing and communication layer includes edge computing servers and V2X communication devices. The edge computing servers are deployed locally in the parking lot to process all the data from the parking lot sensors, generate a global real-time dynamic map, and distribute it to all AVP vehicles. The V2X communication devices can provide a low-latency, high-reliability communication network, such as a 5G base station RSU, to ensure high-speed and stable data transmission between the vehicle, the parking lot, and the cloud.

[0132] Cloud system architecture:

[0133] The cloud platform consists of a scheduling and management system, a high-precision map management system, a user terminal interface, and a monitoring and maintenance platform.

[0134] The dispatch and management system receives parking / retrieval requests from users and performs optimal task allocation based on the location and status of all vehicles, such as assigning the nearest parking space to a newly arrived vehicle. The high-precision map management system stores and distributes high-precision maps of the parking lot, and updates the map data promptly when parking space layouts change or other facilities at the parking lot are updated. The user terminal interface interacts with users via a mobile app. Users can initiate "park" or "retrieve" commands through the app and view the vehicle's location and status in real time. The monitoring and maintenance platform provides parking lot administrators with a platform to monitor the operational status of all AVP vehicles, enabling remote intervention or notification of safety personnel for on-site handling in case of anomalies.

[0135] How does the automatic parking function work?

[0136] The driver drives the car to the parking lot entrance handover area, gets out, and clicks the "Autonomous Parking" command via a mobile app. Upon receiving the request, the cloud platform identifies the vehicle using the parking lot's cameras, assigns it the optimal available parking space, and simultaneously sends the driving route and parking space information to both the vehicle and the parking lot. The driver then locks the car and leaves. After confirming the surrounding environment is safe, the vehicle starts automatically, using its own sensors to perceive the nearby environment and drive into the parking lot. At this time, the parking lot infrastructure transmits a real-time global dynamic map, including the location of moving obstacles, to the vehicle via V2X communication, updating it in real time. The vehicle integrates vehicle-side perception data, parking lot data, and high-precision maps to achieve centimeter-level high-precision positioning. The onboard computing platform makes decisions such as acceleration, deceleration, turning, and obstacle avoidance based on the global path and real-time local environment, executing these decisions through the drive-by-wire system—the entire process is completely autonomous. Upon reaching the assigned parking space, the vehicle uses surround-view cameras and ultrasonic radar for final precise positioning, automatically reverses or drives into the space, and automatically shuts off and locks after coming to a complete stop, sending a "Parking complete, vehicle safe" notification to the user via the app.

[0137] In one embodiment, as shown in Figure 4, a specific implementation process for data fusion is provided, wherein:

[0138] Early fusion is a more advanced and complex multi-sensor fusion strategy in autonomous driving and advanced driver assistance systems. The relationship between early fusion and automatic parking functions can be seen as a "super brain" and "highly coordinated hands." Automatic parking is the "hands," responsible for executing the parking action; early fusion technology is the "super brain," responsible for comprehensively processing all the raw information transmitted from the eyes and ears (sensors), forming the most comprehensive and accurate understanding of the surrounding environment, and then directing the "hands" to complete the action safely and precisely.

[0139] The diagnostic method for vehicle automatic parking function based on sensor front fusion consists of four steps: S1 data acquisition, S2 scene injection, S3 performance assessment, and S4 fault location. The details are as follows:

[0140] S1 Data Acquisition Steps: Synchronously acquire the front fusion layer data of the automatic parking system during operation. This data must include the raw perception data of different sensors, the target list data processed independently by each sensor, and the fused target list data processed by the fusion central processor.

[0141] During testing, this step utilizes a professional data logger to capture multi-layered data from the pre-fusion architecture at high speed and in real-time. This includes raw signals (image frames, echo signals, point clouds) from each sensor in the raw layer (camera, radar, lidar); independent target lists output by each sensor in the target layer after initial processing (vehicle frames identified by the camera, obstacle points detected by the radar); and an authoritative fused target list output by the central processor of the fusion layer after correlating, calibrating, and weighting all the above data according to their trust levels. Obtaining the fused target list is a multi-step, iterative computational process. Its core objective is to integrate the redundant and complementary information provided by multiple heterogeneous sensors into a more accurate, complete, and reliable environmental perception result. Specifically, it includes the following steps:

[0142] S11: Data preprocessing and time synchronization; S12: Target association; S13: Data fusion and state estimation; S14: Trust weighting and conflict resolution; S15: Generate a list of fusion targets.

[0143] S11, data preprocessing and time synchronization, involves unifying the data from all sensors into the same time and spatial coordinate system, including time synchronization and spatial synchronization. Time synchronization refers to using hardware triggers or software timestamps to ensure that all sensors (cameras, radar, lidar) collect data at the same point in time. Spatial synchronization refers to converting the point cloud data of radar and lidar to the pixel coordinate system of camera images through calibration parameters, or, more commonly, unifying all data into a vehicle-centric coordinate system (i.e., X-forward direction, Y-left direction, Z-upward direction).

[0144] S12, Target Association, which determines which targets from different sensor lists actually belong to the same physical object. Commonly used fusion algorithms are the nearest neighbor algorithm and the Hungarian algorithm. The nearest neighbor algorithm calculates the association degree (such as spatial distance, shape similarity, feature similarity) between targets from two sensors (e.g., radar and camera). The Hungarian algorithm (optimal assignment algorithm) finds the globally optimal association pairing among multiple targets, maximizing the overall association degree (or minimizing the total distance).

[0145] S13, Data Fusion and State Estimation, involves merging successfully correlated observations from multiple sensors into a better target state estimate. The target's next position is predicted based on its motion model, and then weighted and averaged with the new sensor observations. The weights depend on the uncertainties of both the prediction and the observations, equivalent to covariance; the smaller the uncertainty, the larger the weight. Logical fusion is then performed on different types of attributes. Cameras provide target classification (e.g., vehicles, pedestrians, bicycles) and color, radar provides accurate distance and velocity, and lidar provides precise 3D contours. The fusion center combines these into a complete target information.

[0146] S14 sets trust weighting, which handles inconsistencies or conflicting information between sensors. The system assigns trust weights to the reliability of different sensors under different operating conditions. In rainy weather, the trust level of the camera may decrease, while the trust level of the radar remains unchanged. When sensor data conflict, such as the camera identifying a "vehicle" and the radar identifying a "guardrail," the fusion algorithm arbitrates based on trust weights, historical data, and contextual information. It typically selects the sensor data with higher trust or outputs a lower overall confidence level.

[0147] S15, Generate the fusion target list, which outputs the final unique fusion target list. After the above steps, the fusion center will generate a unified list. Each target in the fusion target list contains a unique ID, precise 3D position, velocity, size, object type (such as vehicle, pedestrian, bicycle), and existence confidence.

[0148] Based on the specific steps of the diagnostic method described above, here is an example involving the fusion of three types of sensor data: lidar, millimeter-wave radar, and camera data. Consider a scenario where a vehicle is slowly approaching a parallel parking space, and a pedestrian suddenly walks out between two parked vehicles:

[0149] 1) Target layer (each sensor perceives independently):

[0150] The camera detected two parked vehicles (high confidence) and a "pedestrian" silhouette between them, but due to occlusion and lighting, the confidence level is only 60%. Output target:

[0151] [Type: Pedestrian, Location: (X_cam, Y_cam), Confidence: 0.6];

[0152] The millimeter-wave radar detected two parked vehicles (high confidence). A faint, laterally moving dot was detected between the vehicles, but it could not be classified as a pedestrian, small animal, or noise. Output target:

[0153] [Type: Unknown, Location: (X_rad, Y_rad), Lateral velocity: +1.5m / s, Confidence level: 0.5];

[0154] The lidar detected two parked vehicles (high-confidence point cloud clusters). Within the gap between the vehicles, a sparse point cloud cluster approximately 1.7 meters high was captured, resembling the shape of a human leg and torso. Output target:

[0155] [Type: Pedestrian (possible), Location: (X_lidar, Y_lidar), Confidence: 0.7].

[0156] 2) Fusion Layer:

[0157] The first step is coordinate unification and target association. The positions of the three targets are transformed to the vehicle coordinate system. The association algorithm calculates that the three targets are very close in space (distance less than a threshold) and have the same motion trend. Therefore, they are determined to be the same physical object.

[0158] The second step involves data fusion and confidence weighting. First, Kalman filtering is used to fuse the target positions from the three sensors: radar provides accurate distance, lidar provides precise lateral position, and the camera provides pixel-level localization. The fused position (X_fused, Y_fused) is more accurate than data from a single sensor. Next, classification fusion and confidence weighting are performed. The camera and lidar classify the target object as a pedestrian, but their individual confidence levels are low. While the radar cannot classify, its lateral movement speed, provided by the static object (0 velocity) and the vehicle's direction of motion, is strong evidence of a pedestrian. Under good lighting conditions, the fusion algorithm assigns higher classification weights to the camera and lidar, and uses the radar's speed information as strong support. Finally, confidence fusion is performed. Combining information from all three sources, the fusion algorithm calculates a 92% confidence level for the target as a "pedestrian."

[0159] Step 3: Generate the fusion target list. The fusion center outputs the final targets:

[0160] [Unique ID: 103, Type: Pedestrian, Location: (X_fused, Y_fused), Lateral velocity: +1.4m / s, Dimensions: (0.5m, 0.5m, 1.7m), Fusion confidence: 0.92].

[0161] S2 Scenario Injection Step: Inject a preset test scenario into the automatic parking system and record the system's final decision and control output. This step applies a pre-designed, reproducible test scenario to the automatic parking system through simulation tests such as Hardware in the Loop (HIL) testing or real vehicle testing, simulating scenarios such as children "sneaking out" into the parking space.

[0162] S3 Performance Assessment Steps: The fusion target list data described in S1 is precisely compared with the real-world data from the preset test scenario described in S2 to determine whether the overall system performance meets the standards. The key to performance assessment is the quantitative comparison between the fusion target list and the actual ground conditions. Its working principle can be summarized in three key steps: data association, indicator calculation, and performance adjudication.

[0163] 1) Data association:

[0164] Data association is the first step in the comparison process, aiming to determine which target in the fused list is the same physical entity as the object on the ground. The Hungarian algorithm is typically used, aiming to minimize the overall association cost by finding the best match for each ground target. The association cost is the basis for determining whether two targets can be associated, and it depends on spatial distance, overlap, and shape similarity. Spatial distance is calculated by determining the Euclidean distance between the centers of the fused target's bounding box and the ground truth's bounding box; the smaller the distance, the lower the cost. Overlap is calculated by the intersection-over-union ratio (IoU); a higher IoU indicates better overlap and lower cost. Shape similarity is determined by comparing the dimensions (length, width, and height) of the two bounding boxes.

[0165] Simultaneously, an association threshold is set (e.g., a maximum distance of 2 meters or a minimum overlap of 0.5). Only target pairs with costs below the threshold will be successfully associated, meaning the fused target will be successfully paired with a target on the ground. If the association fails, it may be due to a missed detection, where the ground target did not find a matching fused target, meaning the system did not detect this real object; or it may be due to a false alarm, where the fused target did not find a matching ground target, meaning the system fabricated a non-existent object.

[0166] 2) Indicator Calculation:

[0167] Based on data correlation, the index calculation measures system performance through a series of quantifiable indicators, including detection performance indicators, positioning accuracy indicators, and classification accuracy indicators.

[0168] Detection performance metrics include recall, precision, and F1 score.

[0169] The recall rate, calculated as TP / (TP+FN), measures the system's ability to identify real objects. A lower recall rate indicates that more dangerous objects (such as pedestrians and bollards) are missed, resulting in lower system security.

[0170] Accuracy = TP / (TP+FP) is used to measure whether the system's detection results are accurate and reliable. The lower the accuracy, the more false alarms there will be, which will cause the system to frequently perform unnecessary emergency braking ("ghost braking"), affecting comfort and smoothness.

[0171] The F1 score is the harmonic mean of recall and precision, and is a comprehensive indicator of detection performance.

[0172] Positioning accuracy metrics include average translation error and bounding box intersection-over-union ratio (IoU). Average translation error is the average positional error of all TP targets in the X (longitudinal) and Y (lateral) directions. IoU is the average intersection-over-union ratio between all TP targets and their corresponding ground truth bounding boxes; a higher IoU indicates more accurate positioning.

[0173] The classification accuracy metric is the classification accuracy rate, which refers to the proportion of all TP targets (such as vehicles, pedestrians, and bicycles) that are consistent with ground targets.

[0174] 3) Performance Trial:

[0175] The calculated indicators are compared with preset passing thresholds. The decision-making logic is as follows: a "one-vote veto" system is adopted, in which case the overall system performance fails to meet the standard if any key indicator (such as recall rate below 99.9%) fails to meet the standard; or a comprehensive scoring system is adopted, in which different weights are assigned to different indicators, a total score is calculated, and the system meets the standard if the total score is higher than the threshold.

[0176] Based on the above performance evaluation steps, here's an example of a "child peeking out" test scenario: Ground data shows that there is one pedestrian within 5 seconds. The first step is for the system to output a list of fused targets:

[0177] Time T1: No target output (corresponding to actual ground: pedestrians have appeared but are obscured) is not included in the association;

[0178] Time T2: Output target A [Type: pedestrian, Location: P1, Confidence: 0.6] successfully associated, counted as TP;

[0179] Time T3: Output target A [Type: pedestrian, Location: P2, Confidence: 0.8] successfully associated, counted as TP;

[0180] At time T4: Output target B [Type: Vehicle, Location: P3, Confidence: 0.7] has no true ground match, counted as FP (false alarm);

[0181] Time T5: No target output (corresponding to actual ground: pedestrians still exist), counted as FN (missed report).

[0182] Step 2, performance index calculation:

[0183] TP=2 (pedestrians were successfully detected at times T2 and T3), FP=1 (a vehicle was falsely detected at time T4), FN=1 (a pedestrian was missed at time T5).

[0184] The recall rate is 2 / (2+1) = 66.7% (the system only found pedestrians within 66.7% of the time they existed); the precision is 2 / (2+1) = 66.7% (33.3% of the system's detection results are false).

[0185] Step 3, Performance Trial:

[0186] Assume the preset acceptable thresholds are recall > 99% and precision > 95%. The results show a recall of 66.7% < 99% and a precision of 66.7% < 95%, both failing to meet the standards. Therefore, it can be concluded that the system's overall performance in this test scenario is substandard, triggering the subsequent fault location step S4.

[0187] Through the above detailed and quantifiable comparison process, performance judgment is no longer a subjective matter of good or bad, but an objective decision based on data and indicators, providing a clear and reliable basis for subsequent optimization and improvement.

[0188] S4 Fault Location Steps: If the performance does not meet the standard, the system will not directly conclude that it has failed. Instead, it will backtrack and analyze the data of the previous fusion layer layer by layer. By cross-validating and logically reasoning with the data of the fusion layer and the target layer and the original layer, the source of the performance defect will be located and the problem will be accurately determined whether it is caused by a sensor perception error or a decision error of the central fusion algorithm.

[0189] A complex automated parking system can be viewed as a progressive data processing chain consisting of a sensor layer, a target layer, and a fusion layer. When the final output fusion target fails to meet the requirements, the cause is attributed back to the previous link, isolating each layer until the initially faulty link is found. This can be compared to a doctor's diagnosis for fault location; the process is similar. First, the doctor observes a patient's fever (system performance is substandard), then uses blood tests and X-rays (tracing back data from each layer) to determine whether it's a viral infection (sensor malfunction) or a bacterial infection (fusion algorithm defect). Therefore, fault location mainly involves three steps: fusion layer analysis to locate the fusion algorithm problem; target layer analysis to locate the specific sensor problem; and original layer analysis to locate the root cause.

[0190] This section uses the example of a system failing to detect a pedestrian to illustrate the analysis process at each step:

[0191] 1) Analysis of the fusion layer:

[0192] ① Check the consistency of input and output logic. The diagnostic tool will retrieve the target layer data (the target lists output independently by the camera, radar, and lidar) before and after the fault time point and compare and analyze it with the fusion layer data (the final fused target list).

[0193] ② Hypothesis Formulation and Testing. Assume the fusion algorithm incorrectly excluded correct sensor targets. Examine the target lists of each sensor at the time of the failure. At this point, the camera output is: [Pedestrian, confidence 0.7]; the radar output is: [Unknown, with lateral velocity, confidence 0.6]; and the lidar output is: [Pedestrian, confidence 0.65]. All three sensors provided evidence supporting the presence of an object, but the fused target list was empty.

[0194] ③ Locating the root cause of the fault. Based on the above analysis, the problem can be determined to be within the fusion center. Possible causes include: the data association threshold being too strict, requiring the three sensor targets to overlap spatially, while calibration errors lead to association failure; or the trust weighting being unreasonable, with the radar's unknown classification weight set too low, consequently reducing the trustworthiness of other sensors; or the fusion rules having loopholes, only outputting a pedestrian target when all sensor classifications are consistent, with the radar outputting an unknown target causing the rule to be unmet.

[0195] Therefore, the analysis of the fusion layer concludes that the fault can be located in the data association or trust weighting module of the fusion algorithm.

[0196] 2) Target layer analysis:

[0197] The core of target layer analysis lies in determining which sensor perception algorithm gave the wrong answer through cross-validation, hypothesis formulation and testing, and fault localization.

[0198] ① Cross-validation. The fusion layer logic is reasonable, but there is a problem with the input. The diagnostic tool compares each sensor's independent target with the ground target.

[0199] ② Hypothesis formulation and testing. Assume that the perception algorithm of a certain sensor has missed detections or misclassified data. Compare the data at the same time: Ground target: There is a pedestrian; Camera target list output: [Vehicle, confidence 0.9] → Error! (Classification error); Radar target list output: [ ] (empty list) → Error! (Complete missed detection); LiDAR target list output: [Pedestrian, confidence 0.8] → Correct.

[0200] ③ Locating the root cause of the fault. The results showed that the camera misclassified a pedestrian as a vehicle; the radar completely missed detections and did not output any targets. The fusion center received one correct target (LiDAR), one incorrect target (camera), and one missing target (radar). Under these circumstances, even if the fusion algorithm itself is flawless, it is difficult to make a completely correct decision.

[0201] Therefore, the target layer analysis concludes that the fault can be located to a specific sensor, specifically the camera classification algorithm or the radar detection algorithm.

[0202] 3) Original layer analysis:

[0203] The core of primitive layer analysis lies in analyzing why the sensor perception algorithm malfunctions, and whether the problem is caused by hardware or the environment. The specific analysis process includes tracing the data source, hypothesis testing, and hypothesis verification.

[0204] ① To trace the data source, the diagnostic tool needs to retrieve the original layer data at the time of the failure.

[0205] ② Hypothesis and Testing: First, examining the original camera images revealed overexposure or motion blur in the pedestrian areas, resulting in unclear features and misclassification by the classification network. This indicates that the camera algorithm performs poorly under extreme lighting conditions. Next, examining the original radar point cloud revealed that the radar actually received echo points, but these points had low signal-to-noise ratios and were filtered as clutter by the noise filtering module in the radar signal processing. Diagnostic records showed some mud and water on the radar mirror surface. This indicates that sensor contamination caused performance degradation, and the excessively high filtering threshold combined to lead to missed detections. In conclusion, the initial layer concludes that the root cause of sensor failure is significant environmental challenges (lighting and contamination) and insufficient algorithm robustness.

[0206] Thus, through the meticulous process of backtracking and cross-validation described above, fault location achieves the isolation of problems to hardware sensors or software algorithms; the accuracy of pinpointing specific processing steps within the sensor (such as classification, detection, and signal filtering); and the diagnostic conclusions derived from synchronously recorded multi-layered data, making them verifiable and traceable.

[0207] The diagnostic method described above, by synchronously collecting and deeply analyzing multi-layer data in the pre-fusion process, can not only determine whether the system-level performance meets the standards, but also locate the fault to a specific sensor or a certain link in the fusion algorithm. This achieves a leap from "black-box testing" to "white-box diagnosis," greatly shortening the cycle for developers to troubleshoot and fix problems. It is a key technology for improving the reliability and safety of autonomous driving systems.

[0208] The core ideas of fault location reasoning are elimination and root cause analysis.

[0209] Using a process of elimination, we first need to define a range of suspected components. The system's data processing chain is: sensor hardware → raw data → perception algorithm (target layer) → fusion algorithm (fusion layer) → decision control. When parking fails, all components fall within the initial range of suspected components. At this point, we start testing from the end of the data processing chain, checking the first-layer fusion target list. If the fusion target list is correct and complete, accurately containing the location and type of the obstacle ahead, but the vehicle still collides, the problem can be ruled out of the perception system (sensor + fusion), and lies in the decision or control module. However, if the fusion target list itself is incorrect, the fault is locked within the perception system.

[0210] Going back one level, check the independent target lists of each sensor (target layer). If the independent target lists of radar, camera, and lidar are all correct, but the fused result is incorrect, the problem lies in the fusion algorithm, and issues with the sensor hardware and perception algorithm can be ruled out. However, if the fusion result is incorrect, and the independent target of a certain sensor is also incorrect (such as misidentifying a pedestrian as a vehicle), the problem cannot be ruled out outside the sensor layer, and it is necessary to continue going back up.

[0211] Going back one level further, examine the sensor's raw data (raw layer). If the camera's raw image is blurry, overexposed, or obstructed, even the most efficient perception algorithm cannot correctly identify it. In this case, the root cause is insufficient input conditions or hardware failure, and the algorithm itself is ruled out. If the raw image is clear and of high quality, but the target list is still incorrect, the problem can be pinpointed to the specific perception algorithm of the sensor.

[0212] In general, the process of elimination can horizontally locate complex systemic problems to specific problem segments, such as "the fault originates from the classification module of the camera perception algorithm, which malfunctions when the input image quality is normal."

[0213] The core principle of root cause analysis lies in locating the general problem area through a process of elimination, and then delving into the reasons why that area failed in a specific scenario. Its focus is essentially on the causal chain behind the failure. For example, when the problem is "the camera perception algorithm misidentifies pedestrians as vehicles," the root cause analysis process is shown in Figure 5:

[0214] Test Case 1: "Ghost Targets" Appear in the Fusion Results. If there is a target that should not exist in the fusion list, but the check finds that the target exists in the independent list of a certain sensor (e.g., a camera mistakenly identifies a shadow as an obstacle), but other reliable sensors do not detect it, it can be determined that the fusion algorithm over-relies on the erroneous sensor, which is a fusion trust allocation failure.

[0215] Test Case 2: Fusing Results "Missing" Real Targets. If a real target does not appear in the fusion list, then check the individual lists of each sensor one by one.

[0216] If a separate list of all sensors is missing, the scenario may be too extreme.

[0217] If a sensor (such as a camera) detects something, but it is not included in the fusion list, it may be that the fusion algorithm has mistakenly discarded the information, which is a fusion association failure.

[0218] If only one sensor (such as radar) misses detection, while other sensors do, but the fusion list still misses detection, the problem may lie at the fusion level; if the fusion list also misses detection as a result, it can be determined that the failure is due to the failure of that specific sensor (radar).

[0219] Fault modes are categorized into sensor-level and fusion-level. Sensor-level faults include missed detections, false detections, and exceeding detection accuracy limits; fusion-level faults include spatiotemporal synchronization errors, target association errors, and errors in trust level allocation strategies. This categorization provides direct guidance for R&D personnel to quickly understand and fix problems. This method possesses the ability to finely classify diagnostic results, not only locating faults but also labeling them. Furthermore, it has the ability to quantify and statistically analyze diagnostic results. For example, a quantitative indicator can be defined for "camera false detections": the number of false alarms per thousand frames. After multiple tests, this indicator is statistically analyzed, and a final reliability score (e.g., 98.5 points) is assigned to the camera. This transforms performance evaluation from a single "pass / fail" assessment into long-term, traceable reliability growth management.

[0220] The diagnostic methods are applicable to a wide range of data types, including ultrasound, cameras, millimeter-wave radar, and lidar. The methods can be implemented in two ways: on a hardware-in-the-loop (HIL) platform and in real-vehicle testing. For example, the "injecting a preset test scenario" step can be performed on a hardware-in-the-loop (HIL) test platform, where a simulator injects analog signals into the sensor controller and collects the corresponding front-fusion layer data. Alternatively, the "injecting a preset test scenario" step can be performed in a real-vehicle test environment, where a remote-controlled device guides the test target to complete a preset dynamic trajectory, and a high-precision differential GPS system and a third-party reference sensor collect the real-world scene data.

[0221] Compared with the prior art, this application has the following technical advantages:

[0222] 1. By synchronously collecting and comparing data from the original layer, target layer, and fusion layer under the pre-fusion architecture, it is possible not only to determine whether the overall system performance meets the standards, but also to accurately locate the root cause of the fault when the performance does not meet the standards, whether it is a specific sensor (such as a camera or radar) or the fusion algorithm, thus achieving in-depth diagnosis of the automatic parking function from "black box" to "white box".

[0223] 2. By utilizing preset scenario injection and multi-layer data backtracking mechanism, performance judgment and fault location are combined, avoiding the limitation of traditional testing that can only draw a "failure" conclusion. This significantly accelerates the problem investigation and algorithm optimization process, realizes the repeatability and accurate analysis of the testing process, and improves testing efficiency and problem traceability.

[0224] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0225] Based on the same inventive concept, this application also provides an automatic parking function testing device for implementing the automatic parking function testing method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the automatic parking function testing device provided below can be found in the limitations of the automatic parking function testing method described above, and will not be repeated here.

[0226] In an exemplary embodiment, as shown in FIG6, an automatic parking function testing device is provided, including: a data acquisition module 601, a parking test module 602, an index generation module 603, and a backtracking and investigation module 604, wherein:

[0227] Data acquisition module 601 is used to acquire real-world scene data of the vehicle under a preset parking test scenario;

[0228] The parking test module 602 is used to inject a preset parking test scenario into the vehicle's automatic parking system, obtain vehicle perception data collected by the vehicle's on-board sensors during the operation of the automatic parking system, associate target entities based on each vehicle perception data to obtain an independent target entity list, and fuse the vehicle perception data corresponding to each target entity in the independent target entity list to obtain a fused target entity list.

[0229] The indicator generation module 603 is used to obtain the performance indicator data of the automatic parking system based on the fusion target entity list and real scene data.

[0230] The backtracking and troubleshooting module 604 is used to backtrack and troubleshoot the fused target entity list, the independent target entity list, and the perception data of each vehicle step by step when the performance index data indicates that the automatic parking system's performance is not up to standard, in order to find the root cause of the automatic parking system's failure.

[0231] In one embodiment, the automatic parking function testing device described above uses a data acquisition module 601 to acquire real-world scenario data of the vehicle under a preset parking test scenario and transmits the real-world scenario data to an indicator generation module 603. The parking test module 602 then injects the preset parking test scenario into the vehicle's automatic parking system, acquires vehicle perception data collected by the vehicle's onboard sensors during the automatic parking system's operation, and associates target entities based on each vehicle perception data to obtain an independent target entity list. The system then fuses the vehicle perception data corresponding to each target entity in the independent target entity list to obtain a fused target entity list. This fused target entity list is transmitted to the indicator generation module 603. The fused target entity list, the independent target entity list, and each vehicle perception data are then sent to a backtracking and troubleshooting module 604. If the performance indicator data indicates that the automatic parking system's performance is substandard, the backtracking and troubleshooting module 604 performs a step-by-step backtracking and troubleshooting of the fused target entity list, the independent target entity list, and each vehicle perception data to determine the root cause of the automatic parking system's failure. By leveraging a preset scenario injection and a multi-layered data backtracking mechanism, performance assessment is combined with fault location, thereby improving the accuracy of fault location when performance is substandard.

[0232] In one embodiment, the backtracking and investigation module 604 further includes: a quality acquisition submodule, a cause acquisition submodule, and a root cause analysis submodule, wherein:

[0233] The quality acquisition submodule is used to acquire the first data quality of the fused target entity list, the second data quality of the independent target entity list, and the third data quality of the perception data of each vehicle.

[0234] The cause acquisition submodule is used to obtain the cause of the automatic parking system failure based on the first data quality, the second data quality, and the third data quality.

[0235] The root cause analysis submodule is used to perform root cause analysis on the causes of failures and obtain the root causes of failures in the automatic parking system.

[0236] In one embodiment, the cause acquisition submodule is further configured to: determine that the decision or control module of the automatic parking system is faulty if the first data quality meets the standard; determine that the fusion algorithm of the automatic parking system is faulty if the first data quality is substandard but the second data quality meets the standard; and determine the cause of the automatic parking system's fault based on the second data quality and the third data quality if the second data quality is substandard.

[0237] In an exemplary embodiment, the cause acquisition submodule is further configured to: identify faulty vehicle perception data with substandard second data quality from a list of independent target entities; acquire target vehicle perception data of the target vehicle sensor from each vehicle perception data based on the target vehicle sensor corresponding to the faulty vehicle perception data; determine that the target perception algorithm of the target vehicle sensor is faulty if the third data quality of the target vehicle perception data meets the standard; and determine that the input conditions are insufficient or the hardware of the target vehicle sensor is faulty if the third data quality of the target vehicle perception data is substandard.

[0238] In one embodiment, the root cause analysis submodule is further configured to: generate a first preliminary conclusion that the target perception model's generalization ability in a preset parking test scenario is insufficient, and check the model training data of the target perception model; if the check results indicate that the model training data of the target perception model lacks data in the preset parking test scenario, determine that the root cause of the automatic parking system's failure is a defect in the training data; if the check results indicate that the model training data of the target perception model does not lack data in the preset parking test scenario, generate a second preliminary conclusion that the model architecture or algorithm of the target perception model has limitations, and analyze the intermediate layer output of the target perception model; if the analysis results indicate that the target perception model focuses on erroneous features, determine that the root cause of the automatic parking system's failure is a defect in the model attention mechanism; if the analysis results indicate that the target perception model does not focus on erroneous features, generate a third preliminary conclusion that the vehicle perception data preprocessing stage introduces errors, and compare the vehicle perception data and the preprocessed vehicle perception data; if the comparison results indicate that the vehicle perception enhancement parameters are improperly configured, determine that the root cause of the automatic parking system's failure is an error in the data preprocessing pipeline configuration.

[0239] In one embodiment, the indicator generation module 603 is further configured to obtain the correlation degree between each fusion target entity in the fusion target entity list and each physical entity in the real scene data; based on the correlation degree between each fusion target entity and each physical entity, construct a real associated target pair, a first virtual associated target pair, and a second virtual associated target pair; the real associated target pair includes a fusion target entity and a physical entity, the first virtual associated target pair includes a virtual fusion target entity and a physical entity, and the second virtual associated target pair includes a fusion target entity and a virtual physical entity; and obtain the performance indicator data of the automatic parking system according to the real associated target pair, the first virtual associated target pair, and the second virtual associated target pair.

[0240] In an exemplary embodiment, the automatic parking function testing device is further configured to determine that the performance of the automatic parking system is substandard if any performance index data is less than a preset performance index data threshold; or, based on the index weights preset for each performance index data, perform a weighted summation of each performance index data to obtain the overall performance index data of the automatic parking system, and determine that the performance of the automatic parking system is substandard if the overall performance index data is less than a preset overall performance index data threshold.

[0241] In one embodiment, the performance metrics data include recall, precision, F1 score, and classification accuracy. The metric generation module 603 is further configured to obtain the recall of the automatic parking system based on real associated target pairs and a first virtual associated target pair; obtain the precision of the automatic parking system based on real associated target pairs and a second virtual associated target pair; obtain the F1 score of the automatic parking system based on the recall and precision; and obtain the classification accuracy of the automatic parking system based on the classification information of the real associated target pairs and the classification information of the physical entities in the real associated target pairs.

[0242] In one embodiment, the parking test module 602 is further configured to input the vehicle perception data corresponding to each target entity in the independent target entity list into a pre-set Kalman filter for position fusion to obtain the fused position of each target entity; obtain the fusion confidence of each target entity based on the confidence level set for the vehicle sensor; and combine the vehicle perception data, fused position and fusion confidence level corresponding to each target entity to obtain the fused target entity list.

[0243] Each module in the aforementioned automatic parking function testing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of the new energy vehicle in hardware form or independent of it, or stored in the memory of the new energy vehicle in software form, so that the processor can call and execute the corresponding operations of each module.

[0244] In an exemplary embodiment, a new energy vehicle is provided, which can be a server, and its internal structure diagram is shown in Figure 7. The new energy vehicle includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the new energy vehicle provides computing and control capabilities. The memory of the new energy vehicle includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the new energy vehicle stores real-world scene data, vehicle perception data, a list of independent target entities, a list of fused target entities, performance index data, and fault root causes. The I / O interfaces of the new energy vehicle are used for information exchange between the processor and external devices. The communication interface of the new energy vehicle is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an automatic parking function testing method.

[0245] Those skilled in the art will understand that the structure shown in Figure 7 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the new energy vehicle to which the present application is applied. A specific new energy vehicle may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0246] In one exemplary embodiment, a new energy vehicle is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the automatic parking test method of various embodiments.

[0247] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the automatic parking test method of various embodiments.

[0248] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the automatic parking test method of various embodiments.

[0249] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0250] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0251] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0252] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for testing automatic parking function, characterized in that, The method includes: acquiring real-world scenario data of the vehicle under a preset parking test scenario; injecting the preset parking test scenario into the vehicle's automatic parking system, acquiring vehicle perception data collected by the vehicle's onboard sensors during the operation of the automatic parking system, associating target entities based on each of the vehicle perception data to obtain an independent target entity list, and fusing the vehicle perception data corresponding to each target entity in the independent target entity list to obtain a fused target entity list; obtaining performance index data of the automatic parking system based on the fused target entity list and the real-world scenario data; and, if the performance index data indicates that the performance of the automatic parking system is substandard, tracing back the fused target entity list, the independent target entity list, and each of the vehicle perception data level by level to find the root cause of the automatic parking system's failure.

2. The method according to claim 1, characterized in that, The step of tracing back and investigating the fused target entity list, the independent target entity list, and each of the vehicle perception data to obtain the root cause of the automatic parking system's fault includes: obtaining the first data quality of the fused target entity list, the second data quality of the independent target entity list, and the third data quality of each of the vehicle perception data; obtaining the cause of the automatic parking system's fault based on the first data quality, the second data quality, and the third data quality; and performing root cause analysis on the cause of the fault to obtain the root cause of the automatic parking system's fault.

3. The method according to claim 2, characterized in that, The step of obtaining the cause of failure of the automatic parking system based on the first data quality, the second data quality, and the third data quality includes: determining that the decision or control module of the automatic parking system is faulty when the first data quality meets the standard; determining that the fusion algorithm of the automatic parking system is faulty when the first data quality is substandard but the second data quality meets the standard; and determining the cause of failure of the automatic parking system based on the second data quality and the third data quality when the second data quality is substandard.

4. The method according to claim 3, characterized in that, The step of determining the cause of failure in the automatic parking system based on the second data quality and the third data quality includes: identifying faulty vehicle perception data with substandard second data quality from the list of independent target entities; obtaining target vehicle perception data of the target vehicle sensor from each of the vehicle perception data according to the target vehicle sensor corresponding to the faulty vehicle perception data; determining that the target perception algorithm of the target vehicle sensor is faulty if the third data quality of the target vehicle perception data is satisfactory; and determining that the input conditions are insufficient or the hardware of the target vehicle sensor is faulty if the third data quality of the target vehicle perception data is substandard.

5. The method according to claim 2, characterized in that, The root cause analysis of the fault to obtain the root cause of the automatic parking system includes: when the fault is due to a malfunction in the target perception algorithm, generating a first preliminary conclusion that the generalization ability of the target perception model in the preset parking test scenario is insufficient, and checking the model training data of the target perception model; if the check results indicate that the model training data of the target perception model lacks data in the preset parking test scenario, determining that the root cause of the automatic parking system is a training data defect; if the check results indicate that the model training data of the target perception model does not lack data in the parking test scenario, generating the target... The second preliminary conclusion is that the model architecture or algorithm of the perception model has limitations, and the intermediate layer output of the target perception model is analyzed. If the analysis results indicate that the target perception model focuses on erroneous features, the root cause of the automatic parking system failure is determined to be a defect in the model attention mechanism. If the analysis results indicate that the target perception model does not focus on erroneous features, a third preliminary conclusion is generated that the vehicle perception data preprocessing stage introduces errors, and the vehicle perception data and the preprocessed vehicle perception data are compared. If the comparison results indicate that the vehicle perception enhancement parameters are improperly configured, the root cause of the automatic parking system failure is determined to be an error in the data preprocessing pipeline configuration.

6. The method according to claim 1, characterized in that, The step of obtaining the performance index data of the automatic parking system based on the fusion target entity list and the real scene data includes: obtaining the correlation degree between each fusion target entity in the fusion target entity list and each physical entity in the real scene data; constructing real associated target pairs, a first virtual associated target pair, and a second virtual associated target pair based on the correlation degree between each fusion target entity and each physical entity; the real associated target pair includes a fusion target entity and a physical entity, the first virtual associated target pair includes a virtual fusion target entity and a physical entity, and the second virtual associated target pair includes a fusion target entity and a virtual physical entity; and obtaining the performance index data of the automatic parking system based on the real associated target pairs, the first virtual associated target pairs, and the second virtual associated target pairs.

7. The method according to claim 1, characterized in that, The method further includes: for any performance index data, if the performance index data is less than a preset performance index data threshold, determining that the performance of the automatic parking system is substandard; or, based on the index weights preset for each performance index data, performing a weighted summation on each performance index data to obtain the overall performance index data of the automatic parking system, and if the overall performance index data is less than a preset overall performance index data threshold, determining that the performance of the automatic parking system is substandard.

8. The method according to claim 6, characterized in that, The performance metrics data include recall, precision, F1 score, and classification accuracy. Obtaining the performance metrics data of the automatic parking system based on the real-object pairs, the first virtual-object pairs, and the second virtual-object pairs includes: obtaining the recall rate of the automatic parking system based on the real-object pairs and the first virtual-object pairs; obtaining the precision rate of the automatic parking system based on the real-object pairs and the second virtual-object pairs; obtaining the F1 score of the automatic parking system based on the recall rate and the precision rate; and obtaining the classification accuracy rate of the automatic parking system based on the classification information of the real-object pairs and the classification information of the physical entities in the real-object pairs.

9. The method according to claim 1, characterized in that, The step of fusing vehicle perception data corresponding to each target entity in the independent target entity list to obtain a fused target entity list includes: inputting the vehicle perception data corresponding to each target entity in the independent target entity list into a pre-set Kalman filter for position fusion to obtain the fused position of each target entity; obtaining the fusion confidence of each target entity based on the confidence level set for the vehicle sensor; and combining the vehicle perception data, fused position, and fusion confidence of each target entity to obtain the fused target entity list.

10. An automatic parking function testing device, characterized in that, The device includes: a data acquisition module for acquiring real-world scene data of the vehicle under a preset parking test scenario; a parking test module for injecting the preset parking test scenario into the vehicle's automatic parking system, acquiring vehicle perception data collected by the vehicle's onboard sensors during the operation of the automatic parking system, associating target entities based on each of the vehicle perception data to obtain an independent target entity list, and fusing the vehicle perception data corresponding to each target entity in the independent target entity list to obtain a fused target entity list; an indicator generation module for obtaining performance indicator data of the automatic parking system based on the fused target entity list and the real-world scene data; and a backtracking and troubleshooting module for, when the performance indicator data indicates that the automatic parking system's performance is substandard, performing a step-by-step backtracking and troubleshooting of the fused target entity list, the independent target entity list, and each of the vehicle perception data to obtain the root cause of the automatic parking system's failure.

11. A new energy vehicle, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images