An algorithm detection method, computer device and readable storage medium

Abstract

This invention discloses an algorithm detection method, computer device, and readable storage medium, relating to the field of autonomous driving technology. The algorithm is used to detect the perception system of a vehicle with autonomous driving functions, and includes the following steps: acquiring target data from a visual sensor, LiDAR, and millimeter-wave radar respectively, and performing fusion processing to form a fused target, a visual track, and a LiDAR track; calculating the intersection-over-union ratio (IoU) between the target bounding boxes of the visual track and the fused target, and the IoU between the target bounding boxes of the LiDAR track and the fused target for each frame of perceived targets; filtering the perceived targets, and statistically calculating the two IoU ratios corresponding to the filtered perceived targets; and obtaining the detection result based on the statistical calculation results. This invention can determine the impact of the vehicle perception system algorithm on its upstream and downstream processes.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to an algorithm detection method, computer equipment, and readable storage medium. Background Technology

[0002] The process of autonomous driving is essentially a collaborative effort between perception, prediction, and control algorithms to adjust and optimize vehicle hardware. Therefore, in autonomous vehicles, especially in terms of software, each algorithm or strategy is inextricably linked to its upstream and downstream counterparts; a change in one affects the entire system. As infrastructure improves and the scale, quantity, and dimensions of acquired data expand, the algorithms and strategies related to perception fusion are constantly iterating and updating. Since updates to algorithms or strategies impact other upstream and downstream algorithms or strategies, it is crucial to determine whether a new algorithm or strategy will have a positive or negative impact on these other algorithms or strategies before deployment. Current detection technologies largely rely on labeled data or synchronously collected theoretical ground truth data. However, existing detection methods require significant manual labor and time for data labeling, making it difficult to guarantee the reliability of theoretical ground truth. Furthermore, in urban testing scenarios such as NAC (Non-Autonomous Computing) systems, limitations imposed by other related algorithms hinder accurate and consistent detection results. Summary of the Invention

[0003] This invention aims to address, to a certain extent, one of the technical problems in related technologies. To this end, this invention provides an algorithm detection method to determine the impact of a vehicle perception system's algorithm on its upstream and downstream processes.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] An algorithm detection method for detecting the perception system of a vehicle with autonomous driving functions includes the following steps:

[0006] Data on the perceived target are acquired from visual sensors, lidar, and millimeter-wave radar, respectively.

[0007] The data of the perceived target are fused to form a fused target, a visual track, and a lidar track.

[0008] Calculate the first cross-union ratio and the second cross-union ratio of the perceived target data for each frame. The first cross-union ratio is the cross-union ratio between the target box of the visual track and the target box of the fused target, and the second cross-union ratio is the cross-union ratio between the target box of the lidar track and the target box of the fused target.

[0009] The perceived targets are filtered, and the first and second crossover ratios corresponding to the filtered perceived targets are statistically calculated.

[0010] The test results are obtained based on the statistical calculations.

[0011] Optionally, statistical calculations are performed on the first and second intersection-union ratios (IU / U) corresponding to the filtered perceived targets, including:

[0012] For the first intersection-over-union (IoU), calculate the probability distribution, probability density distribution, target box shape distribution, median, mean, standard deviation, and 90th percentile.

[0013] For the second crossover ratio, calculate the probability distribution, probability density distribution, target box distribution, median, mean, standard deviation, and 90th percentile.

[0014] Optionally, obtaining the detection results based on statistical calculations includes the following steps:

[0015] Determine the testing benchmark based on existing data;

[0016] The algorithm being tested is run using existing data. The results are compared with the detection benchmark of the existing data. If the results decrease compared with the detection benchmark, the algorithm being tested has a negative effect on the current vehicle's software and / or hardware. If the results increase compared with the detection benchmark, the algorithm being tested has a positive effect on the current vehicle's software and / or hardware, and the detection benchmark is updated.

[0017] Repeat the following steps until all newly collected data is processed by the algorithm being tested:

[0018] The algorithm being tested is run using the newly acquired data and compared with the updated detection benchmark. If the result decreases compared with the updated detection benchmark, the algorithm being tested has a negative effect on the current vehicle's software and / or hardware; if the result increases compared with the updated detection benchmark, the algorithm being tested has a positive effect on the current vehicle's software and / or hardware, and the detection benchmark is updated.

[0019] Obtain the results of the algorithm being tested running the first cross-union ratio (CUNR), including the 90th percentile, mean, and median, and the results of running the second CUNR. Compare these results with the updated detection benchmark. If the results show an increase compared to the updated benchmark, the algorithm being tested has a positive effect on the current vehicle's software and / or hardware. If any of the results show a decrease compared to the updated benchmark, the algorithm being tested has a negative effect on the current vehicle's software and / or hardware.

[0020] Optionally, obtaining detection results based on statistical calculations also includes:

[0021] Calculate the ratio of the first or second cross-union ratio (CUNR) less than 0.3 to the first or second CUNR of the target bounding box of the fused target whose length or width is less than 1 meter.

[0022] Optionally, obtaining detection results based on statistical calculations also includes addressing the shortcomings of existing algorithm systems for locating perceived targets with a first or second cross-union ratio (CUNR) of less than 0.5, based on a bird's-eye view dynamic map.

[0023] Optionally, the data fusion processing for the perceived target includes the following steps:

[0024] Determine if there are any anomalies in the data acquired from the perceived target and remove them;

[0025] For the first frame of target perception data, a track is started by the lidar. Then, it is determined whether the signal from the visual sensor is related to the track started by the lidar. If they are related, the lidar data is associated. If they are not related, a new track is started by the visual sensor. Then, it is determined whether the track after the millimeter-wave radar data is associated with the lidar data or the track started by the visual sensor is associated. If they are related, the next step is performed. If they are not related, a new track is started.

[0026] After the first frame, the data of the perceived target in each frame is processed in a loop as follows: Kalman filter prediction, LiDAR data association, visual sensor data association, and millimeter-wave radar data association.

[0027] The track output at the end of the loop is the fused target. The target perceived by the lidar is used to form the lidar track through data association, and the target perceived by the visual sensor is used to form the visual track through data association.

[0028] Optionally, the lidar data association involves calculating the weighted average of the track and the lidar's perceived target. The weighted average is the weighted average of the horizontal and vertical distances and the horizontal and vertical velocities of the perceived target, and then the Hungarian matching algorithm is used to pair the two.

[0029] The visual sensor data association involves calculating the weighted average of the trajectory and the perceived target of the visual sensor. The weighted average is the weighted average of the horizontal and vertical distances and the horizontal and vertical velocities of the perceived target. Then, the Hungarian matching algorithm is used to pair the two.

[0030] The millimeter-wave radar data association involves calculating the weighted average of the track and the target sensed by the millimeter-wave radar. The weighted average is the weighted average of the horizontal and vertical distances and the horizontal and vertical velocities of the target sensed. Then, the nearest neighbor pairing algorithm is used to pair the two.

[0031] Optionally, when calculating the first cross-union ratio and the second cross-union ratio of the data of the perceived target in each frame, if the fused target is not associated with the visual sensor in the current frame, the first cross-union ratio is -1; if the fused target is not associated with the lidar in the current frame, the second cross-union ratio is -1.

[0032] Optionally, filtering the perceived target includes the following filtering conditions:

[0033] Targets with a longitudinal position greater than 80 meters or less than -50 meters;

[0034] Targets with a lateral position greater than 8 meters or less than -8 meters;

[0035] The target bounding box of the fused target is a perceived target with a length of less than 1 meter;

[0036] The target bounding box width of the fused target is less than 1 meter;

[0037] The perceived target whose first or second intersection-union ratio is -1.

[0038] Optionally, filtering the perceived target may also include the following filtering conditions:

[0039] For LiDAR-sensing targets, filters are used for targets whose absolute difference in lateral distance between the LiDAR-sensing target and the fused target is greater than 0.8.

[0040] Optionally, the algorithm detection method further includes obtaining relevant information about the fused target, including: timestamp, frame number, sensor ID, horizontal and vertical position, length of the perceived target, width of the perceived target, heading angle, and target category.

[0041] The technical solution provided by this invention, through statistical analysis of parameters such as the intersection-over-union ratio (IoU) in the algorithm, can not only uncover potential performance defects in the perception system and effectively reflect the perception capabilities of the autonomous driving system in real-world complex scenarios, but also accurately determine the impact of the new algorithm on its upstream and downstream systems. Furthermore, based on statistical analysis of parameters such as IoU, data can be verified while detecting the algorithm, achieving closed-loop detection of data and the algorithm.

[0042] Because the technical solution provided by this invention performs statistical analysis on some parameters such as the intersection-union ratio in the algorithm, it greatly reduces labor and time costs compared with the simulation, data labeling or synchronously collected theoretical true data methods used in the prior art.

[0043] Furthermore, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the algorithm detection method described in any one of the above.

[0044] Meanwhile, a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the algorithm detection method described in any one of the above claims.

[0045] These features and advantages of the present invention will be disclosed in detail in the following specific embodiments and accompanying drawings. The preferred embodiments or means of the present invention will be shown in detail in conjunction with the accompanying drawings, but are not intended to limit the technical solutions of the present invention. In addition, each of these features, elements and components appearing in the following text and drawings is a plurality of, and different symbols or numbers are used for convenience of representation, but all represent parts with the same or similar construction or function. Attached Figure Description

[0046] The present invention will be further described below with reference to the accompanying drawings:

[0047] Figure 1 This is a flowchart of an embodiment of the present invention;

[0048] Figure 2 This is an aerial view animation from an embodiment of the present invention. Detailed Implementation

[0049] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described are intended to explain the present invention and should not be construed as limiting the invention.

[0050] The terms "an embodiment," "example," or "trademark" used in this specification refer to a particular feature, structure, or characteristic described in connection with the embodiment itself that may be included in at least one embodiment disclosed in this patent. The phrase "in an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0051] Example:

[0052] This embodiment provides an algorithm detection method for detecting the algorithms of the perception system of a vehicle with autonomous driving capabilities. The perception system of an autonomous vehicle includes visual sensors, LiDAR, and millimeter-wave radar, used to perceive the surrounding environment and provide decision-making basis for autonomous driving. In other words, before a new algorithm applied to the vehicle perception system is put into use, the algorithm detection method provided in this embodiment detects the impact of the algorithm on existing upstream and downstream algorithms or existing hardware.

[0053] like Figure 1 As shown, the algorithm detection method provided in this embodiment includes the following steps:

[0054] Data on the perceived target are acquired from visual sensors, lidar, and millimeter-wave radar, respectively.

[0055] The data from the perceived target are fused to form a fused target, a visual track, and a lidar track. This step includes the following sub-steps:

[0056] The system determines whether the acquired target data is abnormal and removes it. In this embodiment, the target data is acquired from upstream with an abnormality field, so the abnormality of the data can be determined directly through this field.

[0057] During data fusion, the data is always processed sequentially in the order of LiDAR, visual sensor, and millimeter-wave radar. For the first frame of target perception data, a track is started by the LiDAR. Then, it is determined whether the signal from the visual sensor is correlated with the track started by the LiDAR. If they are correlated, the LiDAR data is correlated; otherwise, a new track is started by the visual sensor. Then, it is determined whether the track after the millimeter-wave radar data is correlated with the LiDAR data or the track started by the visual sensor is correlated. If they are correlated, the next step is performed; otherwise, a new track is started.

[0058] Following the first frame, the target data for each frame is processed cyclically as follows: Kalman filter prediction, LiDAR data association, visual sensor data association, and millimeter-wave radar data association. LiDAR data association involves calculating a weighted average of the track and the target data from the LiDAR, where the weighted average is the weighted average of the target's lateral and longitudinal distances and lateral and longitudinal velocities. Then, a Hungarian matching algorithm is used to pair the two. Similarly, visual sensor data association involves calculating a weighted average of the track and the target data from the visual sensor, where the weighted average is the weighted average of the target's lateral and longitudinal distances and lateral and longitudinal velocities. Then, a Hungarian matching algorithm is used to pair the two. Likewise, millimeter-wave radar data association involves calculating a weighted average of the track and the target data from the millimeter-wave radar, where the weighted average is the weighted average of the target's lateral and longitudinal distances and lateral and longitudinal velocities. Then, a nearest neighbor pairing algorithm is used to pair the two. In this embodiment, when calculating the weighted average, the weights are flexibly selected by those skilled in the art based on different vehicle models and configurations; this is a common technique in this field and will not be elaborated further here. The Hungarian matching algorithm and the nearest neighbor matching algorithm are also algorithms with general technical meaning in this field.

[0059] The track output at the end of the loop is the fused target. The target perceived by the lidar is used to form the lidar track through data association, and the target perceived by the visual sensor is used to form the visual track through data association.

[0060] Calculate the first and second cross-union ratios (CUI) of the perceived target data for each frame. The CUI calculation method is a common method in this field, namely:

[0061]

[0062] Here, A and B are two different target bounding boxes. The first cross-union ratio (CUI) is the CUI between the target bounding box of the visual track and the target bounding box of the fused target, and the second CUI is the CUI between the target bounding box of the lidar track and the target bounding box of the fused target. In this embodiment, the target bounding boxes of the visual track and the lidar track are directly obtained from the visual track and the lidar track, respectively, while the target bounding box of the fused target is obtained by the technical personnel through a weighted calculation of the perceived target data from the visual sensor, lidar, and millimeter-wave radar, which form the fused target, according to different vehicle models.

[0063] When calculating the first and second cross-union ratios of the perceived target data in each frame, since the fused target may not be associated with all sensor signals, the first cross-union ratio is -1 if the fused target is not associated with the visual sensor in the current frame, and the second cross-union ratio is -1 if the fused target is not associated with the lidar in the current frame.

[0064] Filter the perceived targets according to the following conditions:

[0065] Targets with a longitudinal position greater than 80 meters or less than -50 meters. Due to limitations in sensor performance, this portion of targets is often inaccurate in actual operation. Therefore, using this portion of targets to measure the performance of the entire sensing system will result in significant bias.

[0066] Targets with a lateral position greater than 8 meters or less than -8 meters are filtered out. Since downstream planning control and target behavior prediction pay little attention to these targets, they have almost no impact on vehicle movement.

[0067] The target bounding box of the fusion target is less than 1 meter in length, and the target bounding box of the fusion target is less than 1 meter in width. Statistical analysis of extensive data reveals that the identification, tracking, and fusion of small targets place high demands on the performance of various sensors. Therefore, when sensor performance is limited, i.e., when a high intersection-union accuracy cannot be achieved, the sensor's output needs continuous monitoring and optimization. This is outside the scope of the technical solution in this embodiment, and therefore, this related information can be temporarily disregarded.

[0068] In addition to the aforementioned filtering conditions, for LiDAR-perceived targets, it is also necessary to filter out targets whose absolute value of the lateral distance difference between the LiDAR-perceived target and the fused target is greater than 0.8. This is because when the intersection-union ratio (IU) between the LiDAR-perceived target and the fused target is less than 0.5, the lateral distance from the vision sensor is prioritized during fusion processing to prevent vehicles from crossing lane lines. Lane line information originates from the vision sensor; if there is a discrepancy between the vision sensor signal and the LiDAR signal, the fused target based on the LiDAR's lateral distance will frequently exhibit lane line violations. This will cause problems for downstream traffic control. Furthermore, to measure the error between LiDAR and the vision sensor, the proportion of data filtered out due to lateral distance differences can be statistically analyzed. Based on the aforementioned filtering conditions, situations limited by sensor capabilities and performance optimization situations that cannot be resolved at the sensor end are identified, effectively reflecting the perception capabilities of the autonomous driving system in real-world complex scenarios.

[0069] In addition to the above conditions, it is also necessary to filter out the sensing targets whose first or second crossover ratio is -1.

[0070] After filtering, statistical calculations are performed on the first and second cross-union ratios (CUI) corresponding to the filtered perceived targets. Specifically, the probability distribution, probability density distribution, target box shape distribution, median, mean, standard deviation, and 90th percentile are calculated for both the first and second CUIs. In this embodiment, the 90th percentile is the 90th percentile whose values ​​are arranged in descending order.

[0071] The test results are obtained based on statistical calculations. This step requires both existing data and newly collected data, and includes the following sub-steps:

[0072] The testing benchmark is determined based on existing data. The testing standard, or benchmark, is formulated by those skilled in the art based on existing algorithms, existing vehicle hardware, and autonomous driving industry standards. This is existing technology and will not be elaborated further here.

[0073] The algorithm under test is run using existing data, and the results are compared with the existing data's detection benchmark. If the results decrease compared to the existing data's detection benchmark, the algorithm under test has a negative effect on the current vehicle's software and / or hardware, indicating that the algorithm or strategy under test may conflict with existing algorithms or existing hardware, and therefore the algorithm under test needs to be adjusted. If the results increase compared to the existing data's detection benchmark, the algorithm under test has a positive effect on the current vehicle's software and / or hardware, and therefore the detection benchmark needs to be updated.

[0074] Repeat the following steps until all newly collected data is processed by the algorithm being tested:

[0075] The algorithm being tested is run using newly acquired data and compared with an updated detection benchmark. If the result decreases compared to the updated benchmark, the algorithm has a negative impact on the vehicle's software and / or hardware; if the result increases compared to the updated benchmark, the algorithm has a positive impact on the vehicle's software and / or hardware, and the benchmark is updated. Therefore, the technical solution provided in this embodiment can verify the data while detecting the algorithm, achieving closed-loop detection of data and the algorithm.

[0076] The results of the tested algorithm running the first intersection-union ratio (IU) and the second IU are obtained, based on the 90th percentile, mean, and median. A higher 90th percentile value indicates a better performance of the autonomous driving perception system. These results are then compared to an updated detection benchmark. If the results show an increase compared to the updated benchmark, the tested algorithm has a positive impact on the vehicle's software and / or hardware. Conversely, if any result shows a decrease compared to the updated benchmark, the tested algorithm has a negative impact on the vehicle's software and / or hardware.

[0077] As can be seen, by statistically analyzing some parameters such as the intersection-union ratio in the algorithm, this embodiment can not only uncover potential performance defects of the perception system and effectively reflect the perception capability of the autonomous driving system in real complex scenarios, but also accurately understand the impact of the new algorithm on its upstream and downstream processes.

[0078] Simultaneously, the ratio of a first or second cross-union ratio (CUNR) less than 0.3 to the first or second CUNR of the target bounding box of the fused target, where the length or width is less than 1 meter, is calculated. Targets with a first or second CUNR less than 0.3 are small targets, and a smaller ratio means more accurate identification of small targets. This further demonstrates that the technical solution provided in this embodiment can effectively reflect the perception capability of the autonomous driving system in real complex scenarios.

[0079] For perceived targets with a first or second intersection-union ratio (IU / R) less than 0.5, the shortcomings of existing algorithm systems for localization based on bird's-eye view dynamic maps are addressed. For example... Figure 2 The track diagram shown is a scatter plot containing the lateral and longitudinal positions, lateral and longitudinal velocities, and lateral and longitudinal accelerations of various sensors and fused targets. Figure 2In the initial frame, the fused target follows the forward millimeter-wave radar signal. However, after frame 2025, the LiDAR, visual sensor, and millimeter-wave radar signals split into signals with different motion states, resulting in low first and second cross-union ratios. This indicates that the visual sensor and LiDAR signals no longer belong to the same object as the millimeter-wave radar signal, and therefore the association should be broken as early as possible to avoid affecting or interfering with the fused target's state update. If the cross-union ratio of the vehicle's fused track is small, it means that the visual sensor signal is now within the fence formed by the millimeter-wave radar signal, indicating a significant deviation in the visual sensor signal. Therefore, visual sensor identification and tracking should be used to address the identification distance issue.

[0080] The algorithm detection method provided in this embodiment also includes a step of acquiring relevant information about the fused target. The acquired relevant information can be used as a filtering condition when analyzing a certain indicator of the detected algorithm. Therefore, this step does not have a specific order in the algorithm detection method provided in this embodiment. The acquired relevant information includes: timestamp, frame number, sensor ID, horizontal and vertical position, length of the sensed target, width of the sensed target, heading angle, and target category.

[0081] Because the technical solution provided in this embodiment performs statistical analysis on some parameters such as the intersection-union ratio in the algorithm, it greatly reduces labor and time costs compared with the simulation, data labeling or synchronously collected theoretical true data methods used in the prior art.

[0082] Meanwhile, this embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described algorithm detection method. The steps of the speed planning method here can be the steps in the memory analysis methods of the various embodiments described above.

[0083] 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. Accordingly, the computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can implement the methods of any of the above embodiments. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0084] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments above. Any modifications that do not depart from the functional and structural principles of the present invention will be included within the scope of the claims.

Claims

1. An algorithm detection method for detecting the perception system of a vehicle with autonomous driving capabilities, characterized in that, The algorithm detection method includes the following steps: Data on the perceived target are acquired from visual sensors, lidar, and millimeter-wave radar, respectively. The data of the perceived target are fused to form a fused target, a visual track, and a lidar track. The first cross-union ratio (CUB) and the second cross-union ratio (CUB) of the target data in each frame are calculated respectively. The first CUB is the cross-union ratio between the target box of the visual track and the target box of the fused target, and the second CUB is the cross-union ratio between the target box of the lidar track and the target box of the fused target. The perceived targets are filtered, and the first and second crossover ratios corresponding to the filtered perceived targets are statistically calculated. The test results are obtained based on statistical calculations. The statistical calculation of the first and second intersection-union ratios corresponding to the filtered perceived targets includes: For the first intersection-over-union (IoU), calculate the probability distribution, probability density distribution, target box shape distribution, median, mean, standard deviation, and 90th percentile. For the second intersection-over-union (IoU), calculate the probability distribution, probability density distribution, target box shape distribution, median, mean, standard deviation, and 90th percentile. Obtaining the test results based on statistical calculations includes the following steps: Determine the testing benchmark based on existing data; The algorithm being tested is run using existing data. The results are compared with the detection benchmark of the existing data. If the results decrease compared with the detection benchmark, the algorithm being tested has a negative effect on the current vehicle's software and / or hardware. If the results increase compared with the detection benchmark, the algorithm being tested has a positive effect on the current vehicle's software and / or hardware, and the detection benchmark is updated. Repeat the following steps until all newly collected data is processed by the algorithm being tested: The algorithm being tested is run using the newly acquired data and compared with the updated detection benchmark. If the result decreases compared with the updated detection benchmark, the algorithm being tested has a negative effect on the current vehicle's software and / or hardware; if the result increases compared with the updated detection benchmark, the algorithm being tested has a positive effect on the current vehicle's software and / or hardware, and the detection benchmark is updated. Obtain the results of the algorithm being tested running the first cross-union ratio (CUNR), including the 90th percentile, mean, and median, and the results of running the second CUNR. Compare these results with the updated detection benchmark. If the results show an increase compared to the updated benchmark, the algorithm being tested has a positive effect on the current vehicle's software and / or hardware. If any of the results show a decrease compared to the updated benchmark, the algorithm being tested has a negative effect on the current vehicle's software and / or hardware.

2. The algorithm detection method according to claim 1, characterized in that, Obtaining test results based on statistical calculations also includes: Calculate the ratio between the first or second cross-union ratio (CUNR) less than 0.3 and the first or second CUNR between the target bounding box of the fused target whose length or width is less than 1 meter.

3. The algorithm detection method according to claim 1, characterized in that, Obtaining detection results based on statistical calculations also includes addressing the shortcomings of existing algorithm systems for locating perceived targets with a first or second cross-union ratio (CUNR) of less than 0.5, based on a bird's-eye view dynamic map.

4. The algorithm detection method according to any one of claims 1 to 3, characterized in that, The data fusion processing for the perceived target includes the following steps: Determine if there are any anomalies in the data acquired from the perceived target and remove them; For the first frame of target perception data, a track is started by the lidar. Then, it is determined whether the signal from the visual sensor is related to the track started by the lidar. If they are related, the lidar data is associated. If they are not related, a new track is started by the visual sensor. Then, it is determined whether the track after the millimeter-wave radar data is associated with the lidar data or the track started by the visual sensor is associated. If they are related, the next step is performed. If they are not related, a new track is started. After the first frame, the data of the perceived target in each frame is processed in a loop as follows: Kalman filter prediction, LiDAR data association, visual sensor data association, and millimeter-wave radar data association. The track output at the end of the loop is the fused target. The target perceived by the lidar is used to form the lidar track through data association, and the target perceived by the visual sensor is used to form the visual track through data association.

5. The algorithm detection method according to claim 4, characterized in that, The lidar data association involves calculating the weighted average of the track and the lidar's perceived target. The weighted average is the weighted average of the horizontal and vertical distances and the horizontal and vertical velocities of the perceived target. Then, the Hungarian matching algorithm is used to pair the two. The visual sensor data association involves calculating the weighted average of the trajectory and the perceived target of the visual sensor. The weighted average is the weighted average of the horizontal and vertical distances and the horizontal and vertical velocities of the perceived target. Then, the Hungarian matching algorithm is used to pair the two. The millimeter-wave radar data association involves calculating the weighted average of the track and the target sensed by the millimeter-wave radar. The weighted average is the weighted average of the horizontal and vertical distances and the horizontal and vertical velocities of the target sensed. Then, the nearest neighbor pairing algorithm is used to pair the two.

6. The algorithm detection method according to claim 4, characterized in that, When calculating the first and second cross-union ratios of the perceived target data in each frame, if the fused target is not associated with the visual sensor in the current frame, the first cross-union ratio is -1; if the fused target is not associated with the lidar in the current frame, the second cross-union ratio is -1.

7. The algorithm detection method according to any one of claims 1 to 3, characterized in that, Filtering the perceived target includes the following filtering conditions: Targets with a longitudinal position greater than 80 meters or less than -50 meters; Targets with a lateral position greater than 8 meters or less than -8 meters; The target bounding box of the fused target is a perceived target with a length of less than 1 meter; The target bounding box width of the fused target is less than 1 meter; The perceived target whose first or second crossover ratio is -1.

8. The algorithm detection method according to claim 7, characterized in that, Filtering the perceived target also includes the following filtering conditions: For LiDAR-sensing targets, filters are used for targets whose absolute difference in lateral distance between the LiDAR-sensing target and the fused target is greater than 0.

8.

9. The algorithm detection method according to any one of claims 1 to 3, characterized in that, The algorithm detection method further includes obtaining relevant information about the fused target, including: timestamp, frame number, sensor ID, horizontal and vertical position, length of the perceived target, width of the perceived target, heading angle, and target category.

10. A computer device 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 algorithm detection method according to any one of claims 1 to 9.

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

Images