Test equivalence method, device and equipment for automatic driving dense pedestrians and vehicle

Abstract

The invention provides a test equivalence method, device and equipment for automatically driving dense pedestrians and a vehicle, the method is applied to the technical field of artificial intelligence and machine learning, and the method comprises the steps: obtaining pedestrian target information sensed by the vehicle based on an automatic driving function; projecting each pedestrian target in the pedestrian target information to a vehicle top view angle to obtain a collision circle; according to the method, the collision circle of each pedestrian target is clustered to obtain a polymer, an equivalent aggregate of dense pedestrians is generated according to the polymer, and the pedestrian perception performance of the automatic driving function is evaluated by using the equivalent aggregate, so that the problems of unreliable pedestrian perception distance measurement index results and the like in dense pedestrian tests in related technologies are solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence and machine learning technology, and specifically to a method, device, equipment and vehicle for testing dense pedestrians in autonomous driving. Background Art

[0002] With the rapid development and progress of the automotive industry and intelligent transportation, as well as the increasingly severe problems such as traffic congestion and frequent traffic accidents caused by the accelerated global urbanization process, people's demand for improving travel efficiency and safety has become more urgent. Therefore, autonomous driving perception technology has come into being. It relies on high-precision sensors, cameras, lidar and other equipment, combined with advanced computer vision, machine algorithms and artificial intelligence algorithms to achieve real-time monitoring and perception of the vehicle's surrounding environment, including the acquisition and processing of information such as vehicles, pedestrians, traffic signs, obstacles, etc., providing key data for route prediction and planning control of autonomous driving vehicles.

[0003] However, related technologies still face many challenges in conducting large-scale truth-value evaluation and verification of perception effects in dense pedestrian scenes. Due to the mutual occlusion between pedestrians and the frequent changes in target positions during movement, it is extremely difficult to establish complete and accurate dense pedestrian truth-value data. At the same time, during the test process, due to the complexity and diversity of pedestrian targets, traditional test matching methods often cannot achieve one-to-one high accuracy and it is difficult to discover key perception problems, which limits the application and development of autonomous driving technology in dense pedestrian scenes. Summary of the invention

[0004] The present application provides an equivalent method, device, equipment and vehicle for testing dense pedestrians in autonomous driving. The method can solve the problem of unreliable results of pedestrian perception ranging indicators in related technical dense pedestrian tests.

[0005] In the first aspect, a test equivalence method for dense pedestrians in autonomous driving is provided, the method comprising: obtaining pedestrian target information perceived by the vehicle based on the autonomous driving function; projecting each pedestrian target in the pedestrian target information to the vehicle's top-down view to obtain a collision circle; clustering the collision circle of each pedestrian target to obtain an aggregate, generating an equivalent aggregate of dense pedestrians based on the aggregate, and using the equivalent aggregate to evaluate the pedestrian perception performance of the autonomous driving function.

[0006] Through the above technical scheme, the embodiment of the present application can obtain the surrounding pedestrian target information in real time and accurately based on the automatic driving function, providing a real data source for subsequent evaluation. After obtaining the pedestrian target information, by projecting the pedestrian target to the top view perspective and generating a collision circle, the representation of the pedestrian target can be simplified, while also retaining the potential collision risk information of the pedestrian to the vehicle, which is helpful for the aggregation of pedestrian targets and the generation of equivalent aggregates in subsequent steps. Then, the aggregates are obtained by clustering the collision circles of each pedestrian target, and further equivalent aggregates are generated, which can simplify the representation of pedestrian targets in the test scene and reduce the complexity of the test data. At the same time, the equivalent aggregate can retain the overall characteristics of dense pedestrians, making the test and evaluation closer to the actual scene, improving the accuracy and effectiveness of the evaluation, and finally using the equivalent aggregate for evaluation, which can more accurately reflect the pedestrian perception ability of the automatic driving function in complex scenes, improve the reliability of the test indicators for perceiving dense pedestrians, and not only reduce the test cost but also improve the test efficiency.

[0007] In combination with the first aspect, in some possible implementation methods, the pedestrian perception performance of the autonomous driving function is evaluated using an equivalent set, including: using the equivalent set to determine the true value data and the data to be tested for each frame; calculating the false detection and missed detection index and the ranging index based on the true value data and the data to be tested for each frame; and evaluating the pedestrian perception performance of the autonomous driving function based on the false detection and missed detection index and the ranging index.

[0008] Through the above technical solution, the embodiment of the present application can merge multiple pedestrian targets in a dense pedestrian scene into a whole for processing through an equivalent aggregate, thereby avoiding missed detection or false detection problems caused by pedestrians occluding or overlapping each other, which helps to improve the accuracy of the evaluation and make the evaluation results more reflective of the true performance of the autonomous driving system. At the same time, the use of equivalent aggregates simplifies the evaluation process, reduces the amount of calculation and time required for the evaluation, and improves the evaluation efficiency. In addition, the embodiment of the present application not only takes into account the false detection and missed detection indicators, but also takes into account the ranging indicators, so that the pedestrian perception performance of the autonomous driving function can be comprehensively evaluated.

[0009] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, the false detection and missed detection indicators and the ranging indicators are calculated based on the true value data and the data to be tested of each frame, including: obtaining a collection of the data to be tested and a collection of the true value data; matching the collection of the data to be tested of the same frame with the collection of the true value data; and calculating the false detection and missed detection indicators and the ranging indicators based on the matched collection.

[0010] Through the above technical scheme, the embodiment of the present application can transform the complex pedestrian perception problem into a matching problem between the sets by obtaining the set of test data and true value data, thereby simplifying the evaluation process. By matching the set of test data and true value data, the correspondence between the pedestrians detected by the system and the actual pedestrians can be accurately found, which provides an accurate basis for the calculation of subsequent false detection and missed detection indicators and ranging indicators. After the matching is completed, by calculating the false detection and missed detection indicators and ranging indicators, the accuracy and reliability of the autonomous driving function in pedestrian perception can be quantified. These indicators can intuitively reflect the performance of the function in the actual test scenario and provide data support for subsequent improvement and optimization.

[0011] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, an aggregate is obtained according to the clustering of the collision circles of each pedestrian target, including: traversing the collision circles of each pedestrian target, randomly selecting the center of the collision circle as the starting point, and connecting all the collision circles within the center target range together as the aggregation result, wherein, if there are no other collision circles within the center target range of the selected collision circle, the selected collision circle will be used as the aggregation result; calculating the center of mass of the aggregation result, cutting off the connection relationship whose distance from the center of mass exceeds the target distance, and using the remaining connection relationship according to the aggregation result as the aggregate.

[0012] Through the above technical scheme, the embodiment of the present application first needs to identify and traverse all detected pedestrian targets, ensuring that all possible pedestrian targets are taken into account, providing a comprehensive data basis for the subsequent aggregation process. Each pedestrian target has a collision circle associated with it. Randomly selecting one from the collision circles as a starting point can increase the robustness of the algorithm and avoid bias against specific pedestrians. At the same time, by connecting the collision circles within the target range, an aggregate representing multiple pedestrian targets can be initially formed, laying the foundation for subsequent precise matching. If there are no other collision circles within the target range of the center of the selected collision circle, the collision circle is used as a separate aggregation result, ensuring that even if there are isolated pedestrian targets, they can be correctly identified and processed, avoiding missed detections. For each aggregation result, by calculating the center of mass and cutting the connection relationship, the aggregate can be further refined to make it more accurately represent the actual pedestrian targets, which helps to reduce false detections and missed detections, improve the accuracy of the evaluation, and finally form an accurate aggregate, providing a reliable basis for subsequent matching with true value data.

[0013] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, an equivalent aggregate of dense pedestrians is generated based on the aggregate, including: obtaining the collision circle score of each collision circle in the aggregate; calculating the weight score of the equivalent aggregate based on the collision circle score of each collision circle; connecting the center of each collision circle in the aggregate, and calculating the center of gravity of the polygonal area formed by the center of the circle; generating an equivalent aggregate of dense pedestrians based on the center of each collision circle in the aggregate, the collision circle score, the weight score and the center of gravity.

[0014] Through the above technical scheme, in the embodiment of the present application, for each collision circle in the aggregate, a score is calculated according to its size, position, occlusion and other factors. The score reflects the visibility and importance of the collision circle in the dense pedestrian scene, which provides basic data for the subsequent weight score calculation, ensuring that the generation of the equivalent aggregate can fully consider the characteristics of each collision circle and improve the accuracy of the evaluation. Then, based on the collision circle score, the weighted average of all collision circle scores in each aggregate is performed to obtain the weight score of the equivalent aggregate, which can more comprehensively evaluate the importance of the equivalent aggregate in the test, provide a strong basis for subsequent matching and index calculation, and help to more accurately reflect the performance of the autonomous driving function in the dense pedestrian scene. By calculating the center of gravity, the subsequent matching and index calculation process is simplified. The center of gravity serves as an important reference point in the matching process, which helps to improve the accuracy and efficiency of the matching. Finally, an equivalent aggregate of dense pedestrians is generated. This equivalent aggregate represents an overall target in the dense pedestrian scene. By converting dense pedestrians into equivalent aggregates, the test process can be simplified and the test efficiency can be improved. At the same time, the equivalent aggregate can more accurately reflect the actual situation in the dense pedestrian scene and improve the reliability and accuracy of the test results.

[0015] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, the collision circle score of each collision circle in the aggregate is obtained, including: obtaining the tangent slope between the coordinate origin of the vehicle coordinate system and each collision circle in the aggregate; forming multiple groups of slope intervals according to the tangent slope of each collision circle, and calculating the occlusion rate of each collision circle according to the multiple groups of slope intervals; calculating the corresponding collision circle score according to the occlusion rate of each collision circle.

[0016] Through the above technical solution, the embodiment of the present application can obtain the relative position and direction relationship between the vehicle and each collision circle by calculating the tangent slope, providing basic data for the subsequent calculation of the occlusion rate and the collision circle score. By calculating the occlusion rate, the visibility of each collision circle in the aggregate can be quantified. The lower the occlusion rate, the easier it is for the collision circle to be blocked by other pedestrians or objects, and thus it is more difficult to be detected in the autonomous driving perception, which provides an important basis for the subsequent calculation of the collision circle score. By calculating the collision circle score, the relative importance of each collision circle in the autonomous driving perception can be quantified. The higher the score of the collision circle, the more it needs to be accurately identified and responded to by the autonomous driving function, which provides an important basis for the perception performance evaluation and testing of the autonomous driving function.

[0017] In the second aspect, a testing equivalent device for dense pedestrians in autonomous driving is provided, which includes: an acquisition module, used to obtain pedestrian target information perceived by the vehicle based on the autonomous driving function; a projection module, used to project each pedestrian target in the pedestrian target information to the vehicle's top view to obtain a collision circle; a testing module, used to cluster the collision circle of each pedestrian target to obtain an aggregate, generate an equivalent aggregate of dense pedestrians based on the aggregate, and use the equivalent aggregate to evaluate the pedestrian perception performance of the autonomous driving function.

[0018] In a third aspect, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-mentioned equivalent methods for testing dense pedestrians in autonomous driving.

[0019] In a fourth aspect, a vehicle is provided, the vehicle having an autonomous driving perception function, wherein the autonomous driving perception function is tested based on any of the above-mentioned autonomous driving dense pedestrian test equivalent methods.

[0020] In a fifth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and the program is executed by a processor to implement any of the above-mentioned equivalent methods for testing dense pedestrians in autonomous driving.

[0021] Beneficial effects of the present invention:

[0022] (1) This application not only simplifies the representation of pedestrian targets in the test scene and reduces the complexity of test data, but also improves the accuracy and effectiveness of the evaluation by obtaining pedestrian target information, projecting to generate collision circles, clustering to generate aggregates and equivalent aggregates, and using equivalent aggregates for evaluation. At the same time, it also helps to reduce test costs and improve test efficiency.

[0023] (2) The present application utilizes an equivalent set to evaluate the pedestrian perception performance of an autonomous driving function, thereby avoiding the problem of false detection or missed detection due to occlusion or overlap, and improving the accuracy of the evaluation. At the same time, the evaluation process is simplified, the amount of calculation and time are reduced, and the evaluation efficiency is improved. The application also comprehensively considers the false detection or missed detection indicators and the ranging indicators, and can comprehensively evaluate the pedestrian perception performance of an autonomous driving function.

[0024] (3) This application calculates the false detection and missed detection index and the ranging index based on the true value data of each frame and the data to be tested. Through the steps of obtaining the aggregate, matching the aggregate and calculating the index, the complex pedestrian perception problem is simplified to the aggregate matching problem, and the corresponding relationship between the data to be tested and the true value data is accurately established, which provides a solid foundation for the calculation of false detection and missed detection and ranging indicators, and further quantifies the accuracy and reliability of autonomous driving in pedestrian perception, and intuitively reflects the test performance.

[0025] (4) This application obtains aggregates based on the collision circles of each pedestrian target, which not only ensures that all pedestrian targets are fully considered, but also significantly improves the accuracy and robustness of pedestrian perception evaluation through random selection, preliminary aggregation, isolated target processing, aggregate refinement, etc., providing a solid foundation for subsequent analysis and optimization.

[0026] (5) The present application generates an equivalent aggregate of dense pedestrians based on the aggregate. By obtaining the score of each collision circle in the aggregate and performing weighted averaging based on the score, the weighted score of the equivalent aggregate is obtained. This not only improves the accuracy of the evaluation, but also simplifies the matching and indicator calculation process by calculating the center of gravity, thereby being able to more accurately reflect the performance of the autonomous driving function in dense pedestrian scenes. The equivalent aggregate finally generated effectively improves the test efficiency and the reliability of the results.

[0027] (6) This application provides an effective method for evaluating the visibility and importance of each collision circle in the aggregate by obtaining the tangent slope, the component slope interval, calculating the occlusion rate and the collision circle score. This method can quantify the occlusion relationship between the collision circles and provide an accurate basis for generating an equivalent aggregate of dense pedestrians. At the same time, the collision circle score can also provide important reference information for subsequent matching and indicator calculation, effectively improving the accuracy and reliability of autonomous driving perception testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a schematic flow chart of an equivalent method for testing dense pedestrians in an autonomous driving system provided in an embodiment of the present application;

[0029] Figure 2 is a schematic diagram of an assembly provided according to an embodiment of the present application;

[0030] Figure 3It is a calculation relationship diagram of EIoU provided according to an embodiment of the present application;

[0031] Figure 4 is a schematic diagram showing the occlusion relationship of a collision circle provided according to an embodiment of the present application;

[0032] Figure 5 is a flowchart of an indicator calculation process of an equivalent method for testing dense pedestrians in an autonomous driving system provided in an embodiment of the present application;

[0033] Figure 6 is a block diagram of an equivalent device for testing dense pedestrians in an autonomous driving system according to an embodiment of the present application;

[0034] Figure 7 It is a schematic diagram of the structure of an electronic device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0035] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0036] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0037] Autonomous driving perception technology can sense the distance, speed and other information of vehicles, pedestrians, cyclists and other targets around the vehicle, so as to predict and plan the route. However, there is no good way to verify the large-scale truth-value evaluation of perception technology for dense pedestrians (such as a group of pedestrians waiting to cross the traffic light at an intersection). The difficulty of its evaluation lies in: the mutual occlusion between pedestrians and the mutual change of target positions during pedestrian movement are very complicated. First, it is difficult to establish complete and accurate truth-value data of dense pedestrians. Second, it is impossible to achieve a one-to-one high accuracy rate for the test matching of dense pedestrians, reach a testable level, and discover key perception problems.

[0038] To this end, the embodiments of the present application propose an equivalence method, device, equipment and vehicle for testing dense pedestrians in autonomous driving. First, in the dense pedestrian scene test of the autonomous driving function, all detected pedestrian targets are identified and traversed, a collision circle is generated for each pedestrian target, and a collision circle score is calculated, which reflects the visibility and importance of the collision circle in the scene. Secondly, according to the scores and positional relationships of the collision circles, the collision circles of multiple pedestrian targets are aggregated to form an equivalent set representing multiple pedestrian targets, and the weight score of the equivalent set is calculated. Then, by calculating the center of gravity of the equivalent set, the subsequent matching with the true value data and the indicator calculation process are simplified. Finally, according to the true value data of each frame and the data to be tested, the false detection and missed detection indicators and the ranging indicators are calculated. This not only improves the reliability of the test indicators for perceiving dense pedestrians, but also provides a data pre-processing method to generate an equivalent set of dense pedestrians, thereby improving the efficiency of indicator testing.

[0039] The following describes the equivalent method, device, equipment and vehicle for testing dense pedestrians in autonomous driving according to the embodiments of the present application with reference to the accompanying drawings.

[0040] Figure 1 It is a schematic flowchart of an equivalent method for testing dense pedestrians in autonomous driving provided in an embodiment of the present application.

[0041] For example, Figure 1 As shown, the equivalent method for testing dense pedestrians in autonomous driving includes:

[0042] In step S101, pedestrian target information perceived by the vehicle based on the automatic driving function is obtained.

[0043] Among them, pedestrian target information refers to the relevant information of surrounding pedestrians obtained by the autonomous driving vehicle through its perception system. This information usually includes the distance information of the pedestrians, that is, the specific position or distance of the pedestrians relative to the autonomous driving vehicle. At the same time, it also includes the speed information of the pedestrians, that is, dynamic data such as the speed or direction of the pedestrians' movement.

[0044] It can be understood that the embodiment of the present application is based on the automatic driving function to obtain the surrounding pedestrian target information in real time and accurately, providing a real data source for subsequent evaluation, thereby improving the reliability of the test results.

[0045] In step S102, each pedestrian target in the pedestrian target information is projected to the top view of the vehicle to obtain a collision circle.

[0046] Among them, the collision circle refers to the circular area obtained by projecting each pedestrian target into the 2D top view of the vehicle. The center of this circle represents the horizontal and vertical distances of the pedestrian target in the vehicle coordinate system. It is used to simulate the spatial range that pedestrians may occupy. It is set according to actual conditions and is not specifically limited.

[0047] It can be understood that the embodiments of the present application can simplify the subsequent data processing steps by converting pedestrian targets into collision circles, making the test evaluation of dense pedestrians more efficient. At the same time, the introduction of collision circles makes the test of pedestrian targets more focused on potential collision risk areas, which helps to more accurately evaluate the performance of autonomous driving vehicles in dense pedestrian scenarios.

[0048] In step S103, a cluster is obtained according to the collision circle clustering of each pedestrian target, an equivalent cluster of dense pedestrians is generated according to the cluster, and the pedestrian perception performance of the autonomous driving function is evaluated using the equivalent cluster.

[0049] Among them, clustering refers to grouping the collision circles of pedestrian targets with similar distances together to form a larger set; aggregate refers to the whole composed of collision circles of multiple pedestrian targets obtained through the clustering process; equivalent aggregate is a virtual entity obtained by further processing based on the aggregate, which is used to evaluate the pedestrian perception performance of the autonomous driving function. It contains the key information of the aggregate, such as the number, position, and score of the collision circles, and is used for subsequent pedestrian perception performance evaluation.

[0050] Pedestrian perception performance refers to the ability of the autonomous driving function to identify, track and understand the surrounding pedestrian targets. In this method, the evaluation of pedestrian perception performance is achieved by comparing the pedestrian targets perceived by the autonomous driving function with the real pedestrian targets (i.e., true value data). By comparing the difference between the two, the perception accuracy, stability and reliability of the autonomous driving function in dense pedestrian scenarios can be evaluated.

[0051] It can be understood that the embodiments of the present application can convert dense pedestrians into units that are easier to handle and evaluate by generating equivalent aggregates, thereby improving the reliability of perceptual test indicators. By clustering dense pedestrians into aggregates and then generating equivalent aggregates, it can effectively reduce the problems of false matches and missed matches caused by mutual occlusion between pedestrians and complex changes in target positions during movement, thereby improving the accuracy of the test. Aggregating multiple pedestrian targets into an equivalent aggregate can reduce the number of matches and the amount of calculation during the test, making the test process more efficient.

[0052] For example, Figure 2 As shown in the figure, the perception result of a single-frame pedestrian is first projected into a coordinate system with the vehicle as the origin to form a collision circle. The center of the circle is the horizontal and vertical distance (x, y) perceived by the target, and the radius is 0.5m. Pedestrian targets that are very close to each other (for example, less than 3m) are aggregated into an equivalent aggregate P1(abcde) (a convex hull polygon with the center of the circle as the vertex), and a single pedestrian with no other pedestrians around it is a separate aggregate P2(f).

[0053] In an embodiment of the present application, an equivalent set is used to evaluate the pedestrian perception performance of an autonomous driving function, including: using the equivalent set to determine the true value data and the data to be tested of each frame; calculating the false detection and missed detection index and the ranging index based on the true value data and the data to be tested of each frame; and evaluating the pedestrian perception performance of the autonomous driving function based on the false detection and missed detection index and the ranging index.

[0054] Among them, true value data refers to data representing the actual pedestrian position and information obtained in actual test scenarios through high-precision sensors or manual labeling; the data to be tested refers to the pedestrian perception results obtained by the autonomous driving function through its own sensors during actual operation or testing; the false detection index refers to the number or proportion of non-pedestrian targets mistakenly identified as pedestrian targets by the autonomous driving function during the perception process; the missed detection index refers to the number or proportion of real pedestrian targets that the autonomous driving function fails to identify during the perception process; the ranging index refers to the accuracy of the autonomous driving function's estimation of the distance to pedestrian targets, which is usually evaluated by comparing the error between the system-estimated distance and the actual distance.

[0055] It can be understood that the embodiments of the present application can combine multiple pedestrian targets in a dense pedestrian scene into a whole for processing through equivalent aggregates, thereby avoiding missed detection or false detection problems caused by pedestrians occluding or overlapping each other, which helps to improve the accuracy of the evaluation and make the evaluation results more reflective of the true performance of the autonomous driving system. At the same time, the use of equivalent aggregates simplifies the evaluation process, reduces the amount of calculation and time required for the evaluation, and improves the evaluation efficiency. In addition, the embodiments of the present application not only consider the false detection and missed detection indicators, but also consider the ranging indicators, so that the pedestrian perception performance of the autonomous driving function can be comprehensively evaluated.

[0056] In an embodiment of the present application, the false detection and missed detection index and the ranging index are calculated based on the true value data and the data to be tested of each frame, including: obtaining a collection of the data to be tested and a collection of the true value data; matching the collection of the data to be tested of the same frame with the collection of the true value data; and calculating the false detection and missed detection index and the ranging index based on the matched collection.

[0057] It can be understood that the embodiments of the present application can transform the complex pedestrian perception problem into a matching problem between the sets by acquiring the set of test data and true value data, thereby simplifying the evaluation process. By matching the set of test data and true value data, the correspondence between the pedestrians detected by the system and the actual pedestrians can be accurately found, which provides an accurate basis for the calculation of subsequent false detection and missed detection indicators and ranging indicators. After the matching is completed, the accuracy and reliability of the autonomous driving function in pedestrian perception can be quantified by calculating the false detection and missed detection indicators and ranging indicators. These indicators can intuitively reflect the performance of the function in the actual test scenario and provide data support for subsequent improvements and optimizations.

[0058] It should be noted that the set of test data of the same frame is matched with the set of true value data, and the Hungarian matching method and EIoU (polygon IoU + centroid deviation + outer rectangle shape length and width deviation) are used as the metric for matching. Figure 3 The calculation relationship diagram of EIoU reduces the cases of missed matches and wrong matches. Due to the generation of the aggregate, the matching threshold can be appropriately enlarged, thereby increasing the number of valid matches and improving the matching accuracy.

[0059] For example, we first use the equivalent set to determine the true value data and the data to be tested for each frame, and calculate the weighted missed detection rate and false detection rate to make the indicator more meaningful. That is, compared with pedestrians exposed on the periphery, the importance of pedestrians surrounded or blocked by other pedestrians (from the perspective of the vehicle) is reduced, to define the algorithm's ability to identify and detect key pedestrians. For scenes without dense pedestrians (the clustered sets are all single targets), the indicator results are the same as traditional indicators.

[0060] Calculation steps for false detection and missed detection indicators:

[0061] 1. The set of perception results (DUT) of the same frame is treated as a target and matched with the set of true values ​​(GT). The matching method uses Hungarian matching and EIoU as the metric:

[0062] EIoU = polygon IoU + centroid deviation + length and width deviation of the outer rectangle shape (EIoU is a general method and will not be introduced in detail here);

[0063] 2. Since a dense collection of pedestrians is generated, the matching threshold can be enlarged to achieve more effective matches and reduce missed matches and wrong matches;

[0064] 3. For the (GT, DUT) pair after matching, define the following basic indicators:

[0065] TP: correct number min (WDUT, WGT);

[0066] FN: missed detection number, when WGT-WDUT is greater than 0, it is: WGT-WDUT, otherwise it is 0;

[0067] FP: number of false positives, when WDUT-WGT is greater than 0, it is: WDUT-WGT, otherwise it is 0;

[0068] 4. Each aggregate is equivalent to a large pedestrian collision body and is counted in the number of missed detections and false detections of this frame;

[0069] 5. Finally, summarize the results of multiple frames to get the final missed detection rate and false detection rate indicators:

[0070] Missed detection rate = FP / (FP+TP)

[0071] False positive rate = FN / (FN+TP).

[0072] The basic idea of ​​calculating the ranging error index is that for pedestrian autonomous driving, more attention is paid to traffic participants facing higher risks. The higher the aggregate weight score W, the more unobstructed pedestrians (traffic participants facing higher risks) the aggregate consists of, so the impact on the error is higher. For scenes without dense pedestrians (clustered aggregates are all single targets), it is the same as the traditional ranging error index.

[0073] The steps for calculating the ranging index are as follows:

[0074] 1. The matching method is consistent with the matching method used in the calculation process of false detection and missed detection indicators;

[0075] 2. For each matching pair, calculate the distance error between their centroids;

[0076] 3. Longitudinal distance error:

[0077] DX=W DUT *|C xDUT -C xGT |

[0078] 4. Lateral distance error:

[0079] DY=W DUT *|C yDUT -C yGT |

[0080] 5. Record the errors of all frames and output them statistically.

[0081] In an embodiment of the present application, an aggregate is obtained by clustering the collision circles of each pedestrian target, including: traversing the collision circles of each pedestrian target, randomly selecting the center of the collision circle as the starting point, and connecting all collision circles within the center target range together as an aggregation result, wherein, if there are no other collision circles within the center target range of the selected collision circle, the selected collision circle will be used as the aggregation result; calculating the center of mass of the aggregation result, cutting off the connection relationships whose distance from the center of mass exceeds the target distance, and using the remaining connection relationships according to the aggregation result as the aggregate.

[0082] It can be understood that the embodiment of the present application first needs to identify and traverse all detected pedestrian targets, ensuring that all possible pedestrian targets are taken into account, providing a comprehensive data basis for the subsequent aggregation process. Each pedestrian target has a collision circle associated with it. Randomly selecting one from the collision circles as a starting point can increase the robustness of the algorithm and avoid bias against specific pedestrians. At the same time, by connecting the collision circles within the target range, an aggregate representing multiple pedestrian targets can be initially formed, laying the foundation for subsequent precise matching. If there are no other collision circles within the target range of the center of the selected collision circle, then the collision circle is used as the aggregation result alone, ensuring that even if there are isolated pedestrian targets, they can be correctly identified and processed, avoiding missed detections. For each aggregation result, by calculating the center of mass and cutting the connection relationship, the aggregate can be further refined to make it more accurately represent the actual pedestrian targets, which helps to reduce false detections and missed detections, improve the accuracy of the evaluation, and finally form an accurate aggregate, providing a reliable basis for subsequent matching with true value data.

[0083] For example, the aggregation method is as follows:

[0084] 1. Randomly select a circle center as the starting point, find the target that is less than 3m away from the point, and connect them together; if there is no point connected to the point, it is regarded as a separate aggregation result;

[0085] 2. Repeat the first step by traversing the centers of the circles connected to it until there are no circles that can be connected;

[0086] 3. Calculate the centroid of the collection and cut off the connections whose distance from the centroid exceeds T meters; (the threshold is selected according to the actual scenario);

[0087] 4. Repeat step 2 until there are no connections that can be cut;

[0088] 5. The obtained aggregate is the aggregate after the final polymerization is completed.

[0089] In an embodiment of the present application, an equivalent aggregate of dense pedestrians is generated based on the aggregate, including: obtaining the collision circle score of each collision circle in the aggregate; calculating the weight score of the equivalent aggregate based on the collision circle score of each collision circle; connecting the center of each collision circle in the aggregate, and calculating the center of gravity of the polygonal area formed by the center of the circle; generating an equivalent aggregate of dense pedestrians based on the center of each collision circle in the aggregate, the collision circle score, the weight score and the center of gravity.

[0090] It can be understood that in the embodiment of the present application, for each collision circle in the aggregate, a score is calculated based on its size, position, occlusion and other factors. The score reflects the visibility and importance of the collision circle in the dense pedestrian scene, which provides basic data for the subsequent weight score calculation, ensuring that the generation of the equivalent aggregate can fully consider the characteristics of each collision circle and improve the accuracy of the evaluation. Then, based on the collision circle score, the weighted average of all collision circle scores in each aggregate is performed to obtain the weight score of the equivalent aggregate, which can more comprehensively evaluate the importance of the equivalent aggregate in the test, provide a strong basis for subsequent matching and index calculation, and help to more accurately reflect the performance of the autonomous driving function in the dense pedestrian scene. By calculating the center of gravity, the subsequent matching and index calculation process is simplified. The center of gravity serves as an important reference point in the matching process, which helps to improve the accuracy and efficiency of the matching. Finally, an equivalent aggregate of dense pedestrians is generated. This equivalent aggregate represents an overall target in the dense pedestrian scene. By converting dense pedestrians into equivalent aggregates, the test process can be simplified and the test efficiency can be improved. At the same time, the equivalent aggregate can more accurately reflect the actual situation in the dense pedestrian scene and improve the reliability and accuracy of the test results.

[0091] For example, Figure 4 As shown, a ray is emitted from the origin O of the vehicle coordinate system and passes through the two tangent points of the collision circle, and N groups of tangent slope intervals [[Ka1,Ka2], [Kb1,Kb2], [Kc1,Kc2], …] are obtained, such as Figure 4 There is a group [Ka1, Ka2] containing another group [Kb1, Kb2], which means that a completely blocks b, that is, when [Ka1, Ka2] ∩ [Kb1, Kb2] = [Kb1, Kb2], the occlusion rate of target b is 100%.

[0092] Find the occlusion rate Nocc of the target N, find the intersection of N and all other slope intervals, and take the group of a targets with the largest intersection interval, where: || means finding the modulus of the interval, that is, the absolute value of Kn1-Kn2. The occlusion rate ranges from [0-1].

[0093] Assign a score S to each collision circle of this collection, S n =1-Nocc .

[0094] For each aggregate, the score of its collision circle is calculated, and the sum of all collision circle scores is used as the weight score W of the aggregate.

[0095] W=∑S n

[0096] Connect the center points of all collision circles of the aggregate into a convex hull polygon area (the general solution is not explained here), and get the convex hull polygon of the dense pedestrian aggregate P. Calculate the centroid point C(x,y) of the aggregate polygon, and get the final equivalent aggregate P containing the following information: (The method of calculating the centroid point of the polygon area is a general method and is not introduced here)

[0097] -The centers of N collision circles [[a x ,a y ],[b x ,b y ]…]

[0098] - The score S of N collision circles a ,S b ,…

[0099] -Total weight score: W (value range: 0 to N)

[0100] -Center of gravity C(C x ,C y )

[0101] According to the above specific embodiments, the process of calculating the false detection, missed detection and ranging index will be described below: Figure 5 As shown in the figure, first, according to the pedestrian target information output by the single-frame perception algorithm, a collision circle is generated, and clustering P is obtained according to the set threshold. Secondly, the occlusion rate of all collision circles in this frame is calculated, and the score Si is obtained. Then, the convex hull polygon is closed for all collision circles of the aggregate, and the polygon center of gravity and the aggregate weight score W are obtained. Finally, the information of this aggregate is used to calculate the index.

[0102] In an embodiment of the present application, a collision circle score of each collision circle in the aggregate is obtained, including: obtaining the tangent slope between the coordinate origin of the vehicle coordinate system and each collision circle in the aggregate; forming multiple groups of slope intervals according to the tangent slope of each collision circle, and calculating the occlusion rate of each collision circle according to the multiple groups of slope intervals; and calculating the corresponding collision circle score according to the occlusion rate of each collision circle.

[0103] It can be understood that the embodiment of the present application can obtain the relative position and direction relationship between the vehicle and each collision circle by calculating the tangent slope, providing basic data for the subsequent calculation of the occlusion rate and the collision circle score. By calculating the occlusion rate, the visibility of each collision circle in the aggregate can be quantified. The lower the occlusion rate, the easier it is for the collision circle to be blocked by other pedestrians or objects, and thus it is more difficult to be detected in the autonomous driving perception, which provides an important basis for the subsequent calculation of the collision circle score. By calculating the collision circle score, the relative importance of each collision circle in the autonomous driving perception can be quantified. The higher the score of the collision circle, the more it needs to be accurately identified and responded to by the autonomous driving function, which provides an important basis for the perception performance evaluation and testing of the autonomous driving function.

[0104] According to the embodiments of the present application, by acquiring pedestrian target information, projecting to generate collision circles, clustering into aggregates and equivalent aggregates, and using equivalent aggregates for evaluation, not only the representation of pedestrian targets in the test scenario is significantly simplified and the complexity of the test data is reduced, but also by quantifying false detection and missed detection indicators, ranging indicators, and visibility and importance scores of collision circles, the reliability and effectiveness of the perception performance evaluation of the autonomous driving function in dense pedestrian scenarios are comprehensively improved. At the same time, by generating equivalent aggregates of dense pedestrians, not only the accuracy of the evaluation is improved, but also the test efficiency and the reliability of the results can be effectively improved.

[0105] Figure 6 It is a block diagram of an equivalent device for testing dense pedestrians in autonomous driving provided in an embodiment of the present application.

[0106] For example, Figure 6 As shown, the autonomous driving dense pedestrian test equivalent device 60 includes: an acquisition module 610, a projection module 620 and a test module 630.

[0107] Among them, the acquisition module 610 is used to obtain pedestrian target information perceived by the vehicle based on the automatic driving function; the projection module 620 is used to project each pedestrian target in the pedestrian target information to the vehicle's top view to obtain a collision circle; the test module 630 is used to cluster the collision circle of each pedestrian target to obtain an aggregate, generate an equivalent aggregate of dense pedestrians based on the aggregate, and use the equivalent aggregate to evaluate the pedestrian perception performance of the automatic driving function.

[0108] It should be noted that the aforementioned explanation of the embodiment of the equivalent method for testing dense pedestrians in autonomous driving is also applicable to the equivalent device for testing dense pedestrians in autonomous driving of this embodiment, and will not be repeated here.

[0109] In summary, the test equivalent device for dense pedestrians in autonomous driving proposed in the embodiment of the present application can capture the pedestrian information around the vehicle in real time and comprehensively by obtaining the pedestrian target information perceived by the vehicle based on the autonomous driving function, so as to provide strong data support for the subsequent real-time evaluation and ensure the accuracy and completeness of the evaluation. By projecting each pedestrian target in the pedestrian target information onto the vehicle's top-down view to obtain a collision circle, the complex three-dimensional spatial information is converted into a two-dimensional top-down view, which greatly reduces the difficulty of analysis and improves the evaluation efficiency. At the same time, the introduction of the collision circle makes the test of pedestrian targets more focused on potential collision risk areas, which helps to more accurately evaluate the performance of autonomous driving vehicles in dense pedestrian scenarios. Finally, the test module can quickly and accurately evaluate the pedestrian perception performance of the autonomous driving function through cluster analysis and the generation of equivalent aggregates, which greatly improves the evaluation efficiency.

[0110] Figure 7 : is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The device may include:

[0111] A memory 701 , a processor 702 , and a computer program stored in the memory 701 and executable on the processor 702 .

[0112] When the processor 702 executes the program, the testing equivalent method for dense pedestrians in autonomous driving provided in the above embodiment is implemented.

[0113] Furthermore, the device also includes:

[0114] The communication interface 707 is used for communication between the memory 701 and the processor 702 .

[0115] The memory 701 is used to store computer programs that can be executed on the processor 702 .

[0116] The memory 701 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0117] If the memory 701, the processor 702 and the communication interface 707 are implemented independently, the communication interface 707, the memory 701 and the processor 702 can be connected to each other through a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0118] Optionally, in a specific implementation, if the memory 701, the processor 702 and the communication interface 707 are integrated on a chip, the memory 701, the processor 702 and the communication interface 707 can communicate with each other through an internal interface.

[0119] The processor 702 may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0120] This embodiment also provides a vehicle having an autonomous driving perception function, wherein the autonomous driving perception function is tested based on any of the above-mentioned autonomous driving dense pedestrian test equivalent methods.

[0121] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement any of the above-mentioned equivalent methods for testing dense pedestrians in autonomous driving.

[0122] Among them, the device, computer-readable storage medium or chip provided in this embodiment is used to execute the corresponding method provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method provided above, and will not be repeated here.

[0123] Through the description of the above implementation methods, technical personnel in the relevant field can understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0124] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0125] The above contents are only specific implementation methods of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A test equivalence method for dense pedestrians in autonomous driving, characterized in that: The following steps are involved: Obtain pedestrian target information perceived by the vehicle based on the autonomous driving function; Projecting each pedestrian target in the pedestrian target information to the top view of the vehicle to obtain a collision circle; An aggregate is obtained according to clustering of collision circles of each pedestrian target, an equivalent aggregate of dense pedestrians is generated according to the aggregate, and the pedestrian perception performance of the autonomous driving function is evaluated using the equivalent aggregate.

2. The test equivalence method for dense pedestrians in autonomous driving according to claim 1, characterized in that: The evaluating the pedestrian perception performance of the autonomous driving function by using the equivalent aggregate includes: Determine the true value data and the data to be tested of each frame by using the equivalent aggregate; Calculating the false detection and missed detection index and the distance measurement index according to the true value data of each frame and the data to be measured; The pedestrian perception performance of the autonomous driving function is evaluated according to the false detection and missed detection index and the ranging index.

3. The test equivalence method for dense pedestrians in autonomous driving according to claim 2, characterized in that: The calculating of the false detection and missed detection index and the ranging index according to the true value data of each frame and the data to be measured includes: Acquire a collection of test data and a collection of true value data; Matching the set of the data to be tested of the same frame with the set of the true value data; The false detection and missed detection index and the distance measurement index are calculated according to the matched set.

4. The test equivalence method for dense pedestrians in autonomous driving according to claim 1, characterized in that: The step of obtaining a cluster according to the collision circle of each pedestrian target includes: Traversing the collision circles of each pedestrian target, taking the center of a randomly selected collision circle as the starting point, connecting all collision circles within the center target range as an aggregation result, wherein if there is no other collision circle within the center target range of the selected collision circle, the selected collision circle is used as the aggregation result; The centroid of the aggregation result is calculated, and the connection relationships whose distance from the centroid exceeds the target distance are cut off, and the remaining connection relationships according to the aggregation result are used as the aggregate.

5. The test equivalence method for dense pedestrians in autonomous driving according to claim 1, characterized in that: The step of generating an equivalent aggregate of dense pedestrians according to the aggregate comprises: Obtaining a collision circle score for each collision circle in the aggregate; Calculating a weight score of the equivalent aggregate according to the collision circle score of each collision circle; Connecting the centers of each collision circle in the aggregate and calculating the center of gravity of the polygonal area formed by the centers of the circles; An equivalent aggregate of the dense pedestrians is generated according to the center of each collision circle in the aggregate, the collision circle score, the weight score and the center of gravity.

6. The test equivalence method for dense pedestrians in autonomous driving according to claim 5, characterized in that: The obtaining of the collision circle score of each collision circle in the aggregate includes: Obtaining the slope of the tangent line between the coordinate origin of the vehicle coordinate system and each collision circle in the aggregate; forming a plurality of groups of slope intervals according to the tangent slopes of each collision circle, and calculating the occlusion rate of each collision circle according to the plurality of groups of slope intervals; The corresponding collision circle score is calculated according to the occlusion rate of each collision circle.

7. A test equivalent device for autonomous driving with dense pedestrians, characterized in that: include: An acquisition module, used to acquire pedestrian target information perceived by the vehicle based on the autonomous driving function; A projection module, used for projecting each pedestrian target in the pedestrian target information to a top view of the vehicle to obtain a collision circle; The test module is used to obtain an aggregate according to the collision circle clustering of each pedestrian target, generate an equivalent aggregate of dense pedestrians according to the aggregate, and use the equivalent aggregate to evaluate the pedestrian perception performance of the autonomous driving function.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the test equivalence method for dense pedestrians in autonomous driving as described in any one of claims 1 to 6.

9. A vehicle, characterized in that: The vehicle has an autonomous driving perception function, wherein the autonomous driving perception function is tested based on the autonomous driving dense pedestrian testing equivalence method described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the testing equivalence method for dense pedestrians in autonomous driving as described in any one of claims 1-6.

Images