Multi-stage matching method, device and computer equipment for automatic driving

Abstract

The application discloses a kind of multi-stage matching method, device and computer equipment for automatic driving.Therein, the method includes: configuring three threshold values including low score threshold, high score threshold and production trajectory threshold, based on the three threshold values, detection frame and trajectory prediction frame are divided into four sets respectively, based on the four sets, by matching high score 3D detection frame with deterministic 3D trajectory prediction frame as the first matching stage, and matching low score 3D detection frame with unmatched deterministic 3D trajectory prediction frame as the second matching stage, and matching unmatched high score 3D detection frame with non-deterministic 3D trajectory prediction frame as the third matching stage, and matching the remaining high score 2D detection frame with the remaining 2D trajectory prediction frame as the fourth matching stage, set two matching standards instead of one matching standard and three threshold values instead of one threshold value.By the above-mentioned mode, the accuracy and robustness of multi-target tracking scheme can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, and in particular to a multi-stage matching method, device and computer equipment for automatic driving. BACKGROUND

[0002] The existing multi-stage matching scheme for automatic driving requires a reliable and robust perception scheme, including a set of sensors and related perception algorithms, as a prerequisite for L4 level automatic driving. In order to meet the safety requirements of automatic driving, it is reasonable to choose a high-precision and information-rich laser radar as the sensor, so it is necessary to design a perception algorithm based on the 3D point cloud output by the laser radar. The algorithm mainly includes two modules, namely a multi-target detection module and a multi-target tracking module. In addition to the accuracy required by the perception scheme, the real-time performance of the vehicle must also be met, that is, the time for the algorithm to process each frame of point cloud must be less than the time for the laser radar to output a frame of point cloud. Since the inference process of the multi-target detection module is time-consuming, the processing time of the multi-target tracking algorithm usually needs to be controlled within a few milliseconds. Therefore, a multi-target tracking algorithm based on Kalman filtering is more common.

[0003] The input of the multi-target tracking algorithm is the 3D detection box output by the multi-target detection module. Each detection box represents a detected object and contains class information such as vehicles, pedestrians, bicycles, etc., position information, size information, etc. Due to the limitations of the detection algorithm itself, there is a possibility of false detection and missed detection, and the detection box does not have the speed information of the object and the inter-frame relationship information. Therefore, the main functions of the tracking algorithm module are as follows:

[0004] 1. Assign the same ID to the same object appearing in the tracking process, which is different from the ID of any other object, and generate a 3D trajectory;

[0005] 2. Filter false detection boxes;

[0006] 3. Supplement the missed detection boxes;

[0007] 4. Supplement the object speed information;

[0008] From this point of view, the tracking algorithm, as a post-processing process of the entire perception scheme, is extremely dependent on the results of the detection module.

[0009] SORT (multi-target tracking) algorithm is a classic multi-target tracking algorithm. This algorithm models the trajectory of each target based on Kalman filtering, updates the state of the corresponding trajectory with each frame of detection box, and accurately associates the current frame of detection box with the predicted trajectory box. The main steps of the matching algorithm of the SORT algorithm and related extension schemes are as follows:

[0010] 1. Calculate the cost matrix between the detection frame and the track prediction frame based on a certain matching standard such as Euclidean distance, 3D IoU (Intersection over Union), etc.

[0011] 2. Use a combination optimization algorithm such as the Hungarian algorithm, greedy algorithm, etc. to find the optimal algorithm combination core parameter as the score threshold of the detection frame. The detection module assigns a score to each detection frame, representing the existence probability of the object and the detection accuracy of the detection frame. Detection frames with low scores generally have no corresponding objects, and detection frames with high scores generally have corresponding objects and indicate that the size and position of the detection frame are accurate. If the set threshold is too low, the perception scheme may output some trajectories without corresponding objects, causing problems such as unnecessary braking; if the threshold is too high, the perception scheme may delay the output of some trajectories, which is not conducive to the decision-making behavior of the vehicle.

[0012] Therefore, it is valuable to study methods that can improve the accuracy of matching algorithms.

[0013] The SORT algorithm has certain shortcomings in the 3D point cloud target matching process:

[0014] 1. Since the matching algorithm described above only matches according to one matching standard, selecting any matching standard will have some related errors, resulting in general accuracy of the multi-target tracking scheme;

[0015] 2. Since the matching algorithm described above only filters low-score detection frames according to one score threshold, it cannot avoid incorrectly filtering some needed detection frames, resulting in general robustness of the multi-target tracking scheme. SUMMARY

[0016] Therefore, the purpose of the present application is to provide a multi-stage matching method, device and computer equipment for automatic driving, which can improve the accuracy and robustness of the multi-target tracking scheme.

[0017] According to one aspect of the present application, a multi-stage matching method for automatic driving is provided, comprising: configuring three thresholds including a low score threshold low , a high score threshold high and a production track threshold spawn; based on the three thresholds, dividing the detection frame and the trajectory prediction frame into four sets respectively; based on the four sets, setting two matching standards instead of one matching standard and setting three thresholds instead of one threshold by matching the high-score 3D detection frame with the determined state 3D trajectory prediction frame as a first matching stage, matching the low-score 3D detection frame with the unmatched determined state 3D trajectory prediction frame as a second matching stage, matching the unmatched high-score 3D detection frame with the non-determined state 3D trajectory prediction frame as a third matching stage, and matching the remaining high-score 2D detection frame with the remaining 2D trajectory prediction frame as a fourth matching stage.

[0018] Wherein, the three thresholds based on the three thresholds, dividing the detection frame and the trajectory prediction frame into four sets respectively, including: based on the three thresholds, using the detection frame with a filtering score less than the low-score threshold α low , dividing the remaining detection frame into low-score detection frame and high-score detection frame according to the high-score threshold β high , and dividing the trajectory prediction frame into determined state trajectory and non-determined state trajectory, and dividing the detection frame and the trajectory prediction frame into four sets respectively.

[0019] Wherein, the high-score 3D detection frame is matched with the determined state 3D trajectory prediction frame as the first matching stage, including: using the highest quality frame matching and matching the easy matching frame first to match the high-score 3D detection frame with the determined state 3D trajectory prediction frame as the first matching stage.

[0020] Wherein, the low-score 3D detection frame is matched with the unmatched determined state 3D trajectory prediction frame as the second matching stage, including: using the low-score detection frame to match the trajectory prediction frame which is not matched in the first matching stage to match the low-score 3D detection frame with the unmatched determined state 3D trajectory prediction frame as the second matching stage.

[0021] Wherein, the remaining high-score 2D detection frame is matched with the remaining 2D trajectory prediction frame as the fourth matching stage, including: using the imaginary image of projecting the remaining frame to the imaginary 2D plane and matching based on the 2D IoU intersection over union to match the remaining high-score 2D detection frame with the remaining 2D trajectory prediction frame as the fourth matching stage.

[0022] According to one aspect of the present application, a multi-stage matching device for automatic driving is provided, comprising: a configuration module, a division module and a setting module; the configuration module is used to configure three thresholds including a low-score threshold α low , a high-score threshold β high and a production trajectory threshold γ spawn; the dividing module is configured to divide the detection boxes and the trajectory prediction boxes into four sets based on the three thresholds; and the setting module is configured to set two matching standards instead of one matching standard and three thresholds instead of one threshold by matching the high-score 3D detection boxes with the certain-state 3D trajectory prediction boxes as a first matching stage, matching the low-score 3D detection boxes with the certain-state 3D trajectory prediction boxes that are not matched as a second matching stage, matching the high-score 3D detection boxes that are not matched with the non-certain-state 3D trajectory prediction boxes as a third matching stage, and matching the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes as a fourth matching stage based on the four sets.

[0023] The dividing module is specifically configured to divide the detection boxes and the trajectory prediction boxes into four sets based on the three thresholds by using the detection boxes with a filtering score less than a low-score threshold α low to divide the remaining detection boxes into low-score detection boxes and high-score detection boxes according to a high-score threshold β high , and divide the trajectory prediction boxes into certain-state trajectories and non-certain-state trajectories.

[0024] The setting module is specifically configured to match the high-score 3D detection boxes with the certain-state 3D trajectory prediction boxes as a first matching stage by using the highest quality boxes for matching and matching the easily matched boxes first.

[0025] The setting module is specifically configured to match the low-score 3D detection boxes with the certain-state 3D trajectory prediction boxes that are not matched as a second matching stage by using the low-score detection boxes to match the trajectory prediction boxes that are not matched in the first matching stage.

[0026] The setting module is specifically configured to match the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes as a fourth matching stage by using the imaginary image of projecting the remaining boxes onto an imaginary 2D plane and matching based on a 2D IoU intersection over union.

[0027] According to still another aspect of the present application, there is provided a computer device, comprising at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the multi-stage matching method for autonomous driving according to any one of the above aspects.

[0028] According to still another aspect of the present application, there is provided a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the multi-stage matching method for autonomous driving according to any one of the above aspects.

[0029] It can be found that the above scheme can configure three thresholds including a low score threshold α low , a high score threshold β high , and a production trajectory threshold γ spawn , and can divide the detection frame and the trajectory prediction frame into four sets based on the three thresholds, and can set two matching standards instead of one matching standard and set three thresholds instead of one threshold in a manner of matching the high-score 3D detection frame with the determined 3D trajectory prediction frame as a first matching stage, matching the low-score 3D detection frame with the unmatched determined 3D trajectory prediction frame as a second matching stage, matching the unmatched high-score 3D detection frame with the non-determined 3D trajectory prediction frame as a third matching stage, and matching the remaining high-score 2D detection frame with the remaining 2D trajectory prediction frame as a fourth matching stage, so as to realize a multi-stage matching algorithm covering more cases by using two matching standards instead of one matching standard and setting three thresholds instead of one threshold, and to improve the accuracy and robustness of the multi-target tracking scheme.

[0030] Further, the above scheme can divide the detection frame and the trajectory prediction frame into four sets in a manner of filtering the detection frame with a score less than the low score threshold α low , dividing the remaining detection frame into low-score detection frames and high-score detection frames according to the high score threshold β high , and dividing the trajectory prediction frame into determined trajectories and non-determined trajectories, which has the advantage of improving the accuracy of dividing the detection frame and the trajectory prediction frame into four sets.

[0031] Further, the above scheme can match the high-score 3D detection frame with the determined 3D trajectory prediction frame as the first matching stage in a manner of matching the highest quality frame and matching the easy-to-match frame first, which has the advantage of reducing the data amount and reducing the difficulty of the following matching stages.

[0032] Further, the above scheme can match the low-score 3D detection frame with the unmatched determined 3D trajectory prediction frame as the second matching stage in a manner of using the low-score detection frame to match the trajectory prediction frame that is not matched in the first matching stage, which has the advantage of checking whether there is a needed low-score frame, and then the unmatched low-score detection frame will not be changed to a new trajectory because there is a threshold of gamma (gamma_spawn) higher than the highest score of the low-score frame.

[0033] Further, the above scheme can adopt the method of projecting the remaining frame to the imaginary image on the imaginary 2D plane and matching based on the 2D IoU intersection ratio to match the remaining high-score 2D detection frame with the remaining 2D track prediction frame as the fourth matching stage, which has the advantage of improving the accuracy of identifying high-speed objects. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0035] Figure 1 is a flow diagram of an embodiment of the multi-stage matching method for autonomous driving of the present application;

[0036] Figure 2 is a structural diagram of an embodiment of the multi-stage matching device for autonomous driving of the present application;

[0037] Figure 3 is a structural diagram of an embodiment of the computer device of the present application. DETAILED DESCRIPTION

[0038] The present application will be further described in detail below in conjunction with the drawings and embodiments. It is particularly pointed out that the following embodiments are only used to illustrate the present application, but do not limit the scope of the present application. Similarly, the following embodiments are only some embodiments of the present application, not all embodiments, and all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0039] The present application provides a multi-stage matching method for autonomous driving, which can improve the accuracy and robustness of the multi-target tracking scheme.

[0040] Please refer to Figure 1 , Figure 1 is a flow diagram of an embodiment of the multi-stage matching method for autonomous driving of the present application. It should be noted that the method of the present application is not limited to the flow order shown in Figure 1 . As shown in Figure 1 , the method comprises the following steps:

[0041] S101: configure three thresholds including a low-score threshold a low , a high-score threshold b high and a production track threshold g spawn .

[0042] In the embodiment, the relationship of the three thresholds can be 0.0≤α low ≤β high ≤γ spawn ≤1.0, or other relationships, which are not limited by the present application.

[0043] S102: Based on the three thresholds, the detection frame and the trajectory prediction frame are respectively divided into four sets.

[0044] The dividing of the detection frame and the trajectory prediction frame into four sets based on the three thresholds can include:

[0045] Based on the three thresholds, the detection frame with a filtering score less than a low score threshold α low is filtered out, the remaining detection frames are divided into low-score detection frames and high-score detection frames according to a high score threshold β high , and the trajectory prediction frame is divided into a certain state trajectory and a non-certain state trajectory, so that the detection frame and the trajectory prediction frame are respectively divided into four sets, which can improve the accuracy of dividing the detection frame and the trajectory prediction frame into four sets.

[0046] S103: Based on the four sets, two matching standards are set instead of one matching standard and three thresholds are set instead of one threshold by matching the high-score 3D detection frame with the certain state 3D trajectory prediction frame as a first matching stage, matching the low-score 3D detection frame with the unmatched certain state 3D trajectory prediction frame as a second matching stage, matching the unmatched high-score 3D detection frame with the non-certain state 3D trajectory prediction frame as a third matching stage, and matching the remaining high-score 2D detection frame with the remaining 2D trajectory prediction frame as a fourth matching stage.

[0047] In the embodiment, the unmatched high-score detection frame is filtered according to a production trajectory threshold γ spawn , and the detection frame below the threshold is filtered out, and the remaining unmatched high-score detection frame and the unmatched trajectory prediction frame containing the certain state and the non-certain state are projected onto a virtual image, and the relationship from the 3D space of the world to the 2D space of the virtual image can be as follows:

[0048]

[0049] In the present embodiment, x is a point in the virtual image 2D space, P is the camera matrix, X is a point in the world 3D space, R is the rotation matrix, t is the translation vector, f is the focal length, p is the principal point offset. K is the calibration matrix, also known as the camera's intrinsic parameters, [R|t] is also known as the camera's extrinsic parameters. R and t are defined as the transformation from the world coordinate to the camera coordinate, the order is rotation first and then translation. In the process of collecting the dataset of the training detection model of the lidar, usually there is a camera and a lidar to record data together, so that the person annotating the data can refer to the video, and this configuration can be used to set the parameters, and finally in the 2D space of the image, the fourth matching stage is carried out.

[0050] In the present embodiment, the matching of the high-score 3D detection frame to the determined state 3D trajectory prediction frame as the first matching stage can include:

[0051] The matching of the high-score 3D detection frame to the determined state 3D trajectory prediction frame as the first matching stage is carried out in the manner of matching the highest quality frame and matching the easily matched frame first, which has the advantage of being able to reduce the amount of data and reduce the difficulty of the subsequent matching stages.

[0052] In the present embodiment, the matching of the low-score 3D detection frame to the determined state 3D trajectory prediction frame that has not been matched as the second matching stage can include:

[0053] The matching of the low-score 3D detection frame to the determined state 3D trajectory prediction frame that has not been matched as the second matching stage is carried out in the manner of using the low-score detection frame to match the trajectory prediction frame that has not been matched in the first matching stage, which has the advantage of being able to see if there are any low-score frames that need to be matched, and then because there is a gamma threshold that is higher than the highest score of the low-score frame, the low-score detection frame that has not been matched will not be changed to a new trajectory.

[0054] In the present embodiment, because the non-determined state trajectory can not have a corresponding object, only their prediction frame can be matched to the high-score detection frame that has not been matched. After the first matching stage, the second matching stage and the third matching stage, most of the detection frames and the corresponding trajectory prediction frames have been matched, and the production trajectory threshold γ spawn After filtering, in most cases all the determined state trajectories can be matched to the corresponding detection frame, and the detection frame of a newly appearing object, for example, has not been matched to the corresponding trajectory. Now only the fast moving objects are left: from the matching point of view, fast moving means that the relative distance between two frames is far, so it can exceed a threshold related to the matching standard and fail to match. Fast moving objects are generally straight, and there are generally no other objects in their recent range, from the perspective of a human driver, the center point of the object does not move too much, and because there are scenes with high speed objects, the object scene is generally not dense, so it is easy to identify as the same object.

[0055] In the fourth matching stage, the remaining high-score 2D detection boxes are matched with the remaining 2D trajectory prediction boxes, which can include:

[0056] In the fourth matching stage, the remaining high-score 2D detection boxes are matched with the remaining 2D trajectory prediction boxes by projecting the remaining boxes onto the imaginary 2D plane and matching based on the 2D IoU, which has the advantage of improving the accuracy of identifying high-speed objects.

[0057] It can be found that in the present embodiment, three thresholds including a low-score threshold α low , a high-score threshold β high , and a production trajectory threshold γ spawn can be configured, and based on the three thresholds, the detection boxes and the trajectory prediction boxes can be divided into four sets, and based on the four sets, two matching standards instead of one matching standard and three thresholds instead of one threshold can be set by matching the high-score 3D detection boxes with the certain-state 3D trajectory prediction boxes in the first matching stage, matching the low-score 3D detection boxes with the unmatched certain-state 3D trajectory prediction boxes in the second matching stage, matching the unmatched high-score 3D detection boxes with the non-certain-state 3D trajectory prediction boxes in the third matching stage, and matching the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes in the fourth matching stage, which can achieve a multi-stage matching algorithm that covers more cases using two matching standards instead of one matching standard and three thresholds instead of one threshold, and can improve the accuracy and robustness of the multi-target tracking scheme.

[0058] Further, in the present embodiment, based on the three thresholds, the detection boxes with a score less than the low-score threshold α low are filtered out, the remaining detection boxes are divided into low-score detection boxes and high-score detection boxes according to the high-score threshold β high , and the trajectory prediction boxes are divided into certain-state trajectories and non-certain-state trajectories, which has the advantage of improving the accuracy of dividing the detection boxes and the trajectory prediction boxes into four sets.

[0059] Further, in the present embodiment, the high-score 3D detection boxes are matched with the certain-state 3D trajectory prediction boxes in the first matching stage by matching the highest quality boxes and matching the easily matched boxes first, which has the advantage of reducing the amount of data and reducing the difficulty of the following matching stages.

[0060] Further, in the embodiment, the low-score 3D detection box can be matched with the non-matched determined state 3D trajectory prediction box in the second matching stage in a manner that the low-score detection box is matched with the trajectory prediction box which fails to match in the first matching stage, and the advantage is that it can be determined whether the low-score box is needed, and the non-matched low-score detection box will not be changed into a new trajectory because there is a gamma threshold higher than the highest score of the low-score box.

[0061] Further, in the embodiment, the remaining high-score 2D detection box can be matched with the remaining 2D trajectory prediction box in the fourth matching stage in a manner that the remaining box is projected onto an imaginary image of a fictitious 2D plane and matched based on a 2D IoU intersection over union, and the advantage is that the accuracy of identifying high-speed objects can be improved.

[0062] The application also provides a multi-stage matching device for autonomous driving, which can improve the accuracy and robustness of the multi-target tracking scheme.

[0063] Please refer to Figure 2 , Figure 2 which is a structural schematic diagram of an embodiment of the multi-stage matching device for autonomous driving. In the embodiment, the multi-stage matching device for autonomous driving 20 comprises a configuration module 21, a division module 22 and a setting module 23.

[0064] The configuration module 21 is configured to configure three thresholds including a low-score threshold α low , a high-score threshold β high and a production trajectory threshold γ spawn .

[0065] The division module 22 is configured to divide the detection box and the trajectory prediction box into four sets based on the three thresholds.

[0066] The setting module 23 is configured to set two matching standards instead of one matching standard and set three thresholds instead of one threshold based on the four sets in a manner that the high-score 3D detection box is matched with the determined state 3D trajectory prediction box in the first matching stage, the low-score 3D detection box is matched with the non-matched determined state 3D trajectory prediction box in the second matching stage, the non-matched high-score 3D detection box is matched with the non-determined state 3D trajectory prediction box in the third matching stage, and the remaining high-score 2D detection box is matched with the remaining 2D trajectory prediction box in the fourth matching stage.

[0067] Optionally, the division module 22 can be specifically configured to:

[0068] divide the detection box and the trajectory prediction box into four sets based on the three thresholds in a manner that the detection box and the trajectory prediction box are filtered based on that the score is less than the low-score threshold α lowthe detection boxes with high scores are divided into high-scored detection boxes and low-scored detection boxes according to a high-score threshold β high The detection boxes and the trajectory prediction boxes are divided into four sets respectively in the manner of dividing the detection boxes into low-scored detection boxes and high-scored detection boxes and dividing the trajectory prediction boxes into certain-state trajectories and non-certain-state trajectories.

[0069] Optionally, the setting module 23 can be specifically used for:

[0070] The high-scored 3D detection boxes are matched with the certain-state 3D trajectory prediction boxes in the first matching stage in the manner of matching the highest-quality boxes and matching the easily-matched boxes first.

[0071] Optionally, the setting module 23 can be specifically used for:

[0072] The low-scored 3D detection boxes are matched with the certain-state 3D trajectory prediction boxes in the second matching stage in the manner of matching the low-scored detection boxes with the trajectory prediction boxes that are not matched in the first matching stage.

[0073] Optionally, the setting module 23 can be specifically used for:

[0074] The high-scored 2D detection boxes that are not matched in the third matching stage are matched with the 2D trajectory prediction boxes that are not matched in the third matching stage in the fourth matching stage in the manner of projecting the remaining boxes onto imaginary images of a fictitious 2D plane and matching based on a 2D IoU intersection over union.

[0075] The various unit modules of the multi-stage matching device 20 for automatic driving can respectively execute the corresponding steps in the above method embodiments, and thus the unit modules will not be described in detail here. For details, please refer to the description of the corresponding steps above.

[0076] The application further provides a computer device, as shown in the accompanying drawings, comprising: at least one processor 31; and a memory 32 communicatively connected with the at least one processor 31; wherein the memory 32 stores instructions executable by the at least one processor 31, and the instructions are executed by the at least one processor 31 to enable the at least one processor 31 to execute the above-described multi-stage matching method for automatic driving. Figure 3

[0077] ​The memory 32 and the processor 31 are connected in a bus manner, the bus can include any number of interconnected buses and bridges, the bus connects one or more processors 31 and various circuits of the memory 32 together. The bus can also connect various other circuits such as peripheral devices, voltage stabilizers and power management circuits together, which are well known in the art, and therefore, further description is not made herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements such as multiple receivers and transmitters, which provide a unit for communicating with various other devices on the transmission medium. The data processed by the processor 31 is transmitted on the wireless medium through the antenna, and further, the antenna also receives data and transmits the data to the processor 31.

[0078] The processor 31 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interface, voltage regulation, power management and other control functions. The memory 32 can be used to store data used by the processor 31 in performing operations.

[0079] The application further provides a computer readable storage medium storing a computer program. The computer program is executed by the processor to implement the above method embodiments.

[0080] It can be found that the above scheme can configure three thresholds including a low score threshold α low , a high score threshold β high , and a production trajectory threshold γ spawn , and based on the three thresholds, the detection frame and the trajectory prediction frame can be divided into four sets respectively, and based on the four sets, two matching standards instead of one matching standard and three thresholds instead of one threshold can be set by matching the high-score 3D detection frame with the determined 3D trajectory prediction frame as the first matching stage, matching the low-score 3D detection frame with the unmatched determined 3D trajectory prediction frame as the second matching stage, matching the unmatched high-score 3D detection frame with the non-determined 3D trajectory prediction frame as the third matching stage, and matching the remaining high-score 2D detection frame with the remaining 2D trajectory prediction frame as the fourth matching stage, so as to realize a multi-stage matching algorithm covering more cases by using two matching standards instead of one matching standard and setting three thresholds instead of one threshold, and improve the accuracy and robustness of the multi-target tracking scheme.

[0081] Further, based on the three thresholds, the above scheme can filter the detection frame with a score less than the low score threshold α low , and the remaining detection frames are matched according to the high score threshold β highThe detection frame and the trajectory prediction frame are divided into four sets in the manner of dividing the detection frame and the trajectory prediction frame into low-score detection frames and high-score detection frames and dividing the trajectory prediction frame into certain-state trajectories and non-certain-state trajectories, and the advantage is that the accuracy of dividing the detection frame and the trajectory prediction frame into four sets can be improved.

[0082] Further, the above scheme can match the high-score 3D detection frame with the certain-state 3D trajectory prediction frame in the first matching stage in the manner of matching the highest-quality frame and matching the easily-matched frame first, and the advantage is that the data amount can be reduced and the difficulty of the following matching stages can be reduced.

[0083] Further, the above scheme can match the low-score 3D detection frame with the non-matched certain-state 3D trajectory prediction frame in the second matching stage in the manner of matching the low-score detection frame with the trajectory prediction frame that is not matched in the first matching stage, and the advantage is that it can be determined whether there is a low-score frame that needs to be matched, and because there is a gamma threshold higher than the highest score of the low-score frame, the low-score detection frame that is not matched will not be changed into a new trajectory.

[0084] Further, the above scheme can match the remaining high-score 2D detection frame with the remaining 2D trajectory prediction frame in the fourth matching stage in the manner of projecting the remaining frame onto an imaginary image of a fictitious 2D plane and matching based on a 2D IoU intersection over union, and the advantage is that the accuracy of identifying high-speed objects can be improved.

[0085] In several embodiments provided by the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the device embodiments described above are merely schematic; the division of the modules or units is merely logical function division; and an actual implementation can be another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0086] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0087] In addition, each function unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software function unit.

[0088] If the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in the form of a contribution to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform all or part of the steps of the methods in each embodiment of the present application. The foregoing storage medium includes various media that can store program codes, such as a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk.

[0089] The above only describes some embodiments of the present application, and does not limit the protection scope of the present application. Any equivalent device or equivalent flow transformation, or direct or indirect application in other related technical fields, which is made by using the content of the present application specification and drawings, is also included in the patent protection scope of the present application.

Claims

1. A multi-stage matching method for automatic driving, characterized in that, The method comprises the following steps: Configuring three thresholds includes a low score threshold , a high score threshold , and a production trajectory threshold ; based on the three thresholds, the detection frame and the trajectory prediction frame are divided into four sets respectively; The four sets are divided by matching high-score 3D detection boxes with certain 3D trajectory prediction boxes as the first matching stage, matching low-score 3D detection boxes with unmatched certain 3D trajectory prediction boxes as the second matching stage, matching unmatched high-score 3D detection boxes with non-certain 3D trajectory prediction boxes as the third matching stage, and matching the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes as the fourth matching stage, setting two matching standards of 3D and 2D instead of one matching standard of 3D, and setting three thresholds instead of the production trajectory threshold One threshold; wherein, The matching of the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes is a fourth matching stage, including: matching the high-score detection boxes that are not matched according to a production trajectory threshold The matching of the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes is a fourth matching stage, including: matching the high-score detection boxes that are not matched according to a production trajectory threshold 2. The multi-stage matching method for automatic driving of claim 1, wherein, The method based on the three thresholds, the detection frame and the trajectory prediction frame are divided into four sets, comprising: Based on the three threshold values, the detection boxes with a filtering score less than a low score threshold value are filtered out The remaining detection boxes are divided into low-score detection boxes and high-score detection boxes according to a high score threshold value The trajectory prediction boxes are divided into certain-state trajectories and non-certain-state trajectories. In this way, the detection boxes and the trajectory prediction boxes are divided into four sets respectively. 3.The multi-stage matching method for automatic driving of claim 1, wherein, The first matching stage of matching the high-score 3D detection frame with the determined state 3D trajectory prediction frame comprises: The first matching stage of matching the high-score 3D detection frame with the determined state 3D trajectory prediction frame comprises: 4.The multi-stage matching method for automatic driving of claim 1, wherein, The second matching stage of matching the low-score 3D detection frame with the determined state 3D trajectory prediction frame which is not matched comprises: The second matching stage of matching the low-score 3D detection frame with the determined state 3D trajectory prediction frame which is not matched comprises:

5. A multi-stage matching device for automatic driving, characterized by, The method comprises the following steps: The method comprises the following steps: The configuration module is configured to configure three thresholds including a low score threshold , a high score threshold , and a production trajectory threshold . The division module is used for dividing the detection frame and the trajectory prediction frame into four sets based on the three thresholds; The setting module is used for setting two matching standards of 3D and 2D instead of one matching standard of 3D and setting three threshold values instead of the production trajectory threshold value by dividing in the following manners: matching high-score 3D detection boxes with determined state 3D trajectory prediction boxes as a first matching stage, matching low-score 3D detection boxes with determined state 3D trajectory prediction boxes that are not matched as a second matching stage, matching high-score 3D detection boxes that are not matched with non-determined state 3D trajectory prediction boxes as a third matching stage, and matching the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes as a fourth matching stage. one threshold value; wherein, The matching of the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes is a fourth matching stage, including: matching the high-score detection boxes that are not matched according to a production trajectory threshold The matching of the remaining high-score 2D detection boxes with the remaining 2D trajectory prediction boxes is a fourth matching stage, including: matching the high-score detection boxes that are not matched according to a production trajectory threshold 6. The multi-stage matching apparatus for automatic driving according to claim 5, wherein The division module is specifically used for: Based on the three thresholds, the detection boxes with a filtering score less than a low score threshold are filtered out The remaining detection boxes are divided into low-score detection boxes and high-score detection boxes according to a high score threshold The trajectory prediction boxes are divided into certain-state trajectories and non-certain-state trajectories. In this way, the detection boxes and the trajectory prediction boxes are divided into four sets respectively.

7. The multi-stage matching apparatus for automatic driving according to claim 5, wherein The setting module is specifically used for: the first matching stage of matching the high-score 3D detection frame with the determined state 3D trajectory prediction frame comprises:

8. A computer device, comprising: The method comprises the following steps: At least one processor; And a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the multi-stage matching method for automatic driving according to any one of claims 1 to 4.

9. A computer-readable storage medium, characterized in that, The computer program is stored in the memory, and when the computer program is executed by the processor, the multi-stage matching method for automatic driving according to any one of claims 1 to 4 is realized.

Images