Multi-target tracking method and system based on sparse attention mechanism and storage medium
By constructing a 3D target tracking method using a sparse attention mechanism, the problems of high computational complexity and low real-time performance in existing technologies are solved, achieving efficient and accurate 3D multi-target tracking and improving robustness and speed.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2024-01-26
- Publication Date
- 2026-07-07
AI Technical Summary
Existing 3D multi-target tracking methods have high computational complexity and low real-time performance. Furthermore, traditional global tracking methods have limited effectiveness within the image plane and cannot effectively handle motion state estimation of 3D targets.
By employing a sparse attention mechanism, a sparse attention matrix is constructed between the target features at the current time and those at historical time, reducing the attention computation for irrelevant targets, achieving global association matching, and improving computational efficiency and accuracy.
It effectively reduces the problems of false association and missed targets, improves the accuracy and robustness of association, and significantly reduces computational complexity, thereby increasing the processing speed of the tracking algorithm.
Smart Images
Figure CN117994775B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving environmental perception technology, and in particular to a multi-target tracking method, system and storage medium based on a sparse attention mechanism. Background Technology
[0002] In recent years, with the continuous development of autonomous driving technology, environmental perception using point cloud data acquired by LiDAR has gradually become a crucial aspect of ensuring the safety of autonomous vehicles. Building upon accurate 3D target detection based on techniques such as PointRCNN, CenterPoint, and PointPillars, how to further combine temporal multi-frame observations to estimate the target's motion state has become a research hotspot in the perception field in recent years. Traditional 3D multi-target tracking methods mostly employ local tracking, i.e., pre-setting the target's temporal trajectory, then gradually adding new observations based on Hungarian matching and Kalman filtering methods, and updating the trajectory accordingly, such as AB3DMOT.
[0003] Currently, the field of visual 2D multi-object tracking has seen the emergence of global tracking, which has achieved more robust tracking results. However, its application scenarios are limited, and it can only track targets within the image plane. At the same time, as the introduced time length increases, the computational complexity of the attention module increases quadratically, resulting in large computational delays and low real-time performance of the algorithm. Summary of the Invention
[0004] To address the aforementioned problems, the purpose of this invention is to provide a multi-target tracking method, system, and storage medium based on a sparse attention mechanism, which achieves faster processing speed, higher accuracy, and greater robustness in three-dimensional multi-target tracking.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: a multi-target tracking method based on a sparse attention mechanism, comprising: performing three-dimensional target detection on acquired multi-frame laser point cloud data, obtaining the target detection result at the current time and the target detection result at the historical time respectively, and encoding the target features at the current time and the target features at the historical time respectively; constructing an attention matrix between the target features at the current time and the target features at the historical time, calculating an attention mask between the target at the current time and the target at the historical time, and obtaining a sparse attention matrix between the target at the current time and the global targets at all historical times through the attention mask; performing weighted processing on all target features at the historical time according to the sparse attention matrix to obtain a state estimation result; and obtaining the association results at all historical times according to the sparse attention matrix at different times in the historical frame, and using the association results at all historical times as the final association matching result.
[0006] Furthermore, the acquired multi-frame laser point cloud data is used for 3D target detection, including:
[0007] A three-dimensional target detector is used to perform three-dimensional target detection on the laser point cloud data. The detection result is the three-dimensional spatial position information of multiple targets at the current time, which is used as the target detection result at the current time.
[0008] The acquired historical multi-frame laser point cloud data are processed by a 3D target detector to obtain the historical detection results. All historical detection results are then arranged together as the historical target detection results.
[0009] Furthermore, target feature encoding is performed on the target detection results at the current moment, including:
[0010] The target detection result at the current moment is feature encoded. The encoder consists of two stacked multilayer sensing mechanisms, and the encoder outputs the target features at the current moment.
[0011] Furthermore, target feature encoding is performed on the target detection results at historical moments, including:
[0012] The target detection results at historical time are feature encoded by using two stacked MLPs with different parameters as encoders, which output the target features at historical time.
[0013] Furthermore, a sparse attention matrix is obtained through an attention mask between the current target and the global target at all historical time points, including:
[0014] Assuming the attention mask is And initialize the elements therein to 1, representing that there needs to be a correlation between the target at any current moment and the target at any historical moment;
[0015] Calculate any target at the current time and any target at any historical moment The three-dimensional spatial distance d between them ij Arrange them together according to the target index to obtain the distance matrix.
[0016] Set distance threshold d th If the distance between two targets is greater than the threshold d th If the two targets do not require attention calculation, then attention calculation is considered unnecessary; otherwise, attention calculation is required, and the attention mask between the current target and the historical target is obtained based on the attention calculation.
[0017] Based on the obtained attention mask, calculate the sparse attention matrix between the current target and the global target at all historical time steps.
[0018] Furthermore, the target features at all historical moments are weighted according to the sparse attention matrix, including:
[0019] The sparse attention matrix is used as the weight matrix to weight all target features at historical time points, resulting in the optimized current time-point state features.
[0020] Based on the optimized current state characteristics, the state estimation head is calculated using a two-layer MLP to obtain the state estimation result.
[0021] Furthermore, obtaining the final association matching results includes:
[0022] At different times in the historical frames, the sparse attention matrix is decomposed into multiple sub-attention matrices between the current time target and the historical time targets;
[0023] After normalizing the sub-attention matrix and taking the maximum index, the target with the highest correlation with the target at the current time in different historical time periods is obtained as the correlation result;
[0024] The final association matching result is obtained by piecing together the association results from all historical moments.
[0025] A multi-target tracking system based on a sparse attention mechanism includes: a target detection and target feature encoding module, which performs 3D target detection on acquired multi-frame laser point cloud data, obtains the target detection result at the current time and the target detection result at the historical time, and encodes the target features at the current time and the target features at the historical time respectively; a sparse attention module, which constructs an attention matrix between the target features at the current time and the target features at the historical time, calculates an attention mask between the target at the current time and the target at the historical time, and obtains a sparse attention matrix between the target at the current time and the global targets at all historical time through the attention mask; and a state estimation result and association matching result output module, which performs weighted processing on all target features at the historical time according to the sparse attention matrix to obtain the state estimation result; and obtains the association results at all historical time according to the sparse attention matrix at different times of the historical frame, and uses the association results at all historical time as the final association matching result.
[0026] A computer-readable storage medium storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform any of the methods described above.
[0027] A computing device includes: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.
[0028] The present invention has the following advantages due to the adoption of the above technical solutions:
[0029] 1. This invention adopts the idea of global association matching, which effectively reduces the problems of false association and missed targets caused by upstream detection errors, and greatly improves the accuracy and robustness of association.
[0030] 2. In the process of global feature interaction, the present invention designs a sparse attention mechanism, which effectively filters out attention calculations between irrelevant targets in the scene. While ensuring that the correlation effect is not reduced, the computational complexity is greatly reduced and the processing speed of the tracking algorithm is improved. Attached Figure Description
[0031] Figure 1 This is a flowchart of a multi-target tracking method based on sparse attention mechanism in an embodiment of the present invention;
[0032] Figure 2 This is a flowchart of the target association result output module in an embodiment of the present invention. Detailed Implementation
[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.
[0034] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0035] In one embodiment of the present invention, a multi-target tracking method based on a sparse attention mechanism is provided. In this embodiment, as... Figure 1 As shown, the method includes the following steps:
[0036] 1) After performing 3D target detection on the acquired multi-frame laser point cloud data, the target detection results at the current time and the target detection results at the historical time are obtained respectively, and the target features are encoded respectively to obtain the target features at the current time and the target features at the historical time;
[0037] 2) Construct the attention matrix between the target features at the current time and the target features at the historical time, calculate the attention mask between the target at the current time and the target at the historical time, and obtain the sparse attention matrix between the target at the current time and the global target at all historical time through the attention mask;
[0038] 3) After weighting all target features at historical moments according to the sparse attention matrix, the state estimation result is obtained; and at different times of the historical frame, the association results of all historical moments are obtained according to the sparse attention matrix, and the association results of all historical moments are used as the final association matching result.
[0039] In this embodiment, the input information for the tracker is the result of three-dimensional target detection.
[0040] In step 1) above, the acquired multi-frame laser point cloud data is used for 3D target detection, including the following steps:
[0041] 1.1) A three-dimensional target detector is used to perform three-dimensional target detection on the laser point cloud data. The detection result is the three-dimensional spatial position information of multiple targets at the current time, which is used as the target detection result at the current time.
[0042] For example, for laser point cloud data acquired at the current time t, the publicly available 3D target detector PointPillars is used for 3D target detection. The detection result is the 3D spatial position information of multiple targets at the current time, which is represented in the form of bounding boxes in this embodiment, i.e., the target detection result at time t. Where N t Let be the number of targets at time t. The attributes representing the i-th target include the target's three-dimensional spatial location. Three-dimensional size yaw angle direction and its semantic category
[0043] 1.2) The acquired historical multi-frame laser point cloud data are processed by the three-dimensional target detector to obtain the historical detection results, and all the historical detection results are arranged together as the historical target detection results.
[0044] For example, in this embodiment, ten historical point clouds are used as input, denoted as P. τ P τ-1 ,. ..., P τ-9 After being processed by a 3D target detector, the target detection result F is obtained. τ F τ-1 F τ-9 In this embodiment, all detection results from historical moments are arranged together to obtain:
[0045]
[0046] It contains N h =N τ-1 +N τ-2 +…+N τ-9 One goal.
[0047] Then F τ and F h As input to the global tracker, where F τ The shape is N τ ×7, F h The shape is N h ×7.
[0048] In step 1) above, the target feature encoding consists of two stages. The first stage is: to encode the target feature of the target detection result at the current time. Specifically, the encoder is composed of two stacked multilayer sensing mechanisms, and the encoder outputs the target feature at the current time.
[0049] In this embodiment, the first stage is to target F at the current time. τ Feature encoding is performed, a process consisting of two stacked multi-layer perceptron (MLP) layers. The specific computation process is as follows:
[0050]
[0051] in, This represents the intermediate features output by the first layer of the perceptron during the feature encoding stage. Here are the weights and biases of the first layer MLP, and σ(·) is the activation function. In this embodiment, σ(·) = ReLU(·) is used as the activation function. The weights and biases of the second-layer MLP are C1 = C1 = 128 in this embodiment. This is the output of the model, i.e., the target features at the current time.
[0052] In step 1) above, the second stage is: to encode the target features of the target detection results at historical time. Specifically, the target detection results at historical time are encoded using two stacked MLPs with two different sets of parameters as encoders, and the encoders output the target features at historical time.
[0053] In this embodiment, the second stage involves historical time F. h Feature encoding is performed, which uses two stacked MLPs with different parameters as encoders. The same computation process as in the first stage is employed to obtain... and This is the output of the encoder, i.e., the target features at historical moments.
[0054] In step 2) above, based on the target feature encoding, the feature representation of the target at the current time is obtained. And two sets of expressions of target characteristics at historical moments. and Based on this, the attention matrix between the two can be calculated using the following formula:
[0055] A = Softmax(Q × K)
[0056] Where Softmax(·) is the normalization function. It represents the magnitude of attention between the current target Q and the historical target K.
[0057] In this embodiment, the sparse attention matrix between the current target and the global target at all historical moments is obtained through an attention mask, including the following steps:
[0058] 2.1) Assume the attention mask is And initialize the elements therein to 1, representing that there needs to be a correlation between the target at any current moment and the target at any historical moment;
[0059] 2.2) Calculate any target at the current time. and any target at any historical moment The three-dimensional spatial distance d between them ij Arrange them together according to the target index to obtain the distance matrix.
[0060] 2.3) Set the distance threshold d th If the distance between two targets is greater than the threshold d th If the two targets do not require attention calculation, then attention calculation is considered unnecessary; otherwise, attention calculation is required, and the attention mask between the current target and the historical target is obtained based on the attention calculation.
[0061] Optional, distance threshold dth =30m.
[0062] Optionally, attention calculation yields an attention mask. When the attention mask between target i and target j is True, set M ij =1, otherwise set M ij =0. Therefore, the attention mask between the current target and the historical target is obtained.
[0063] 2.4) Based on the obtained attention mask, calculate the sparse attention matrix between the current target and the global target at all historical time steps.
[0064] In step 3) above, based on the sparse attention matrix The weighted processing of all target features at historical moments includes the following steps:
[0065] 3.1.1) The sparse attention matrix As a weight matrix, it applies to all target features at historical moments. We perform weighted analysis to obtain the optimized current state features.
[0066] Optimized current state features:
[0067]
[0068] 3.1.2) Based on the optimized current state characteristics, the state estimation head is calculated using a two-layer MLP to obtain the state estimation result.
[0069] The two-layer MLP calculation is as follows:
[0070] H r =σ(Q) r W (r1) +b (r1) )
[0071] S r =H r W (r2) +b (r2)
[0072] In the formula, This represents the intermediate features of the first layer MLP output in the state estimation head. These are the weights and biases of the first layer MLP, and σ(·) is the activation function. In this embodiment, σ(·) = ReLU(·) is also used as the activation function. For the weights and biases of the second-layer MLP, this embodiment uses C.r1 =64, C r2 =12, which is the final output. C r2 =12 represents the state estimation result for each target, corresponding to N τ This represents the target number at the current moment. Where, [x] i y, z i [h] indicates the location of the target. i w i , l i ] represents the size of the target, r i Indicates the yaw angle of the target, c i Indicates the semantic category of the target. This indicates the target's velocity in the top-down view. This indicates the target's acceleration as seen in the top-down view.
[0073] In step 3) above, if Figure 2 As shown, obtaining the final association matching result includes the following steps:
[0074] 3.2.1) At different times in the historical frames, the sparse attention matrix is... It is decomposed into multiple sub-attention matrices between the current time-instance target and the historical time-different targets;
[0075] For example, when obtaining a sparse attention matrix Based on this, and according to the different moments in which the historical frames are located, they can be broken down into... in This represents the attention matrix between the current target and the target at time t.
[0076] 3.2.2) After normalizing the sub-attention matrix and taking the maximum value index, the target with the highest correlation with the target at the current time from different historical time points is obtained as the correlation result. This establishes a correlation between the current time objective and the time objective at time t:
[0077]
[0078] 3.2.3) By concatenating the association results from all historical moments, the final association matching result A is obtained. asso :
[0079]
[0080] It represents the correlation between all targets at the current moment and all targets at any historical moment.
[0081] In one embodiment of the present invention, a multi-target tracking system based on a sparse attention mechanism is provided, comprising:
[0082] The target detection and target feature encoding module performs 3D target detection on the acquired multi-frame laser point cloud data, obtains the target detection results at the current time and the target detection results at the historical time, and encodes the target features at the current time and the target features at the historical time respectively.
[0083] The sparse attention module constructs an attention matrix between the target features at the current time and the target features at historical time, calculates the attention mask between the target at the current time and the target at historical time, and obtains the sparse attention matrix between the target at the current time and the global target at all historical time through the attention mask;
[0084] The state estimation and association matching output module calculates the state estimation result by weighting all target features at historical time points based on the sparse attention matrix. It then calculates the association results for all historical time points based on the sparse attention matrix at different times in the historical frame and uses these association results as the final association matching result.
[0085] In the above embodiments, the acquisition of multiple frames of laser point cloud data is used for three-dimensional target detection, including:
[0086] A three-dimensional target detector is used to perform three-dimensional target detection on the laser point cloud data. The detection result is the three-dimensional spatial position information of multiple targets at the current time, which is used as the target detection result at the current time.
[0087] The acquired historical multi-frame laser point cloud data are processed by a 3D target detector to obtain the historical detection results. All historical detection results are then arranged together as the historical target detection results.
[0088] In the above embodiments, target feature encoding is performed on the target detection result at the current moment, including:
[0089] The target detection result at the current moment is feature encoded. The encoder consists of two stacked multilayer sensing mechanisms, and the encoder outputs the target features at the current moment.
[0090] In the above embodiments, target feature encoding is performed on the target detection results at historical time points, including:
[0091] The target detection results at historical time are feature encoded by using two stacked MLPs with different parameters as encoders, which output the target features at historical time.
[0092] In the above embodiments, obtaining the sparse attention matrix between the current target and the global target at all historical time points through an attention mask includes:
[0093] Assuming the attention mask is And initialize the elements therein to 1, representing that there needs to be a correlation between the target at any current moment and the target at any historical moment;
[0094] Calculate any target at the current time and any target at any historical moment The three-dimensional spatial distance d between them ij Arrange them together according to the target index to obtain the distance matrix.
[0095] Set distance threshold d th If the distance between two targets is greater than the threshold d th If the two targets do not require attention calculation, then attention calculation is considered unnecessary; otherwise, attention calculation is required, and the attention mask between the current target and the historical target is obtained based on the attention calculation.
[0096] Based on the obtained attention mask, calculate the sparse attention matrix between the current target and the global target at all historical time steps.
[0097] In the above embodiments, the weighted processing of all target features at historical time points based on the sparse attention matrix includes:
[0098] The sparse attention matrix is used as the weight matrix to weight all target features at historical time points, resulting in the optimized current time-point state features.
[0099] Based on the optimized current state characteristics, the state estimation head is calculated using a two-layer MLP to obtain the state estimation result.
[0100] In the above embodiments, obtaining the final association matching result includes:
[0101] At different times in the historical frames, the sparse attention matrix is decomposed into multiple sub-attention matrices between the current time target and the historical time targets;
[0102] After normalizing the sub-attention matrix and taking the maximum index, the target with the highest correlation with the target at the current time in different historical time periods is obtained as the correlation result;
[0103] The final association matching result is obtained by piecing together the association results from all historical moments.
[0104] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.
[0105] In one embodiment of the present invention, a computing device is provided, which can be a terminal and may include: a processor, a communication interface, memory, a display screen, and an input device. The processor, communication interface, and memory communicate with each other via a communication bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs, which, when executed by the processor, implement the methods described in the above embodiments. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, a management network, NFC (Near Field Communication), or other technologies. The display screen can be a liquid crystal display or an e-ink display. The input device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad mounted on the casing of the computing device, or an external keyboard, touchpad, or mouse. The processor can call logical instructions stored in the memory.
[0106] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0107] In one embodiment of the present invention, a computer program product is provided, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to perform the methods provided in the above-described method embodiments.
[0108] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, which stores server instructions that cause a computer to perform the methods provided in the above embodiments.
[0109] The computer-readable storage medium provided in the above embodiments has a similar implementation principle and technical effect to the above method embodiments, and will not be described again here.
[0110] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0111] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0112] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A multi-target tracking method based on sparse attention mechanism, characterized in that, include: After performing 3D target detection on the acquired multi-frame laser point cloud data, the target detection results at the current time and the target detection results at the historical time are obtained respectively, and the target features are encoded respectively to obtain the target features at the current time and the target features at the historical time. Construct an attention matrix between the target features at the current time step and the target features at historical time steps, calculate the attention mask between the target at the current time step and the target at historical time steps, and obtain a sparse attention matrix between the target at the current time step and the global target at all historical time steps through the attention mask, including: assuming the attention mask is... It initializes its elements to 1, signifying that there needs to be a correlation between the target at any current moment and the target at any historical moment; it calculates the target at any current moment. and any target at any historical moment Three-dimensional spatial distance between Arrange them together according to the target index to obtain the distance matrix. Set distance threshold If the distance between two targets is greater than a threshold If the two targets do not require attention calculation, then attention calculation is considered unnecessary; otherwise, attention calculation is required, and the attention mask between the current target and the historical target is obtained based on the attention calculation. Based on the obtained attention mask, calculate the sparse attention matrix between the current target and the global target at all historical time steps. ; After weighting all target features at historical moments using the sparse attention matrix, the state estimation result is obtained; and at different times in the historical frame, the association results at all historical moments are obtained based on the sparse attention matrix, and the association results at all historical moments are used as the final association matching result. The sparse attention matrix is used to weight all target features at historical time points, including: using the sparse attention matrix as a weight matrix to weight all target features at historical time points to obtain optimized current time-point state features; and using a two-layer MLP to calculate the state estimation head based on the optimized current time-point state features to obtain the state estimation result.
2. The multi-target tracking method based on sparse attention mechanism as described in claim 1, characterized in that, The acquired multi-frame laser point cloud data is used for 3D target detection, including: A three-dimensional target detector is used to perform three-dimensional target detection on the laser point cloud data. The detection result is the three-dimensional spatial position information of multiple targets at the current time, which is used as the target detection result at the current time. The acquired historical multi-frame laser point cloud data are processed by a 3D target detector to obtain the historical moment detection results. All historical moment detection results are then arranged together as the historical moment target detection results.
3. The multi-target tracking method based on sparse attention mechanism as described in claim 1, characterized in that, The target feature is encoded based on the target detection result at the current moment, including: The target detection result at the current moment is feature encoded. The encoder consists of two stacked multilayer sensing mechanisms, and the encoder outputs the target features at the current moment.
4. The multi-target tracking method based on sparse attention mechanism as described in claim 1, characterized in that, Target feature encoding is performed on the target detection results at historical moments, including: The target detection results at historical time are feature encoded by using two stacked MLPs with different parameters as encoders, which output the target features at historical time.
5. The multi-target tracking method based on sparse attention mechanism as described in claim 1, characterized in that, The final acquisition of the association matching results includes: At different times in the historical frames, the sparse attention matrix is decomposed into multiple sub-attention matrices between the current time target and the historical time targets; After normalizing the sub-attention matrix and taking the maximum index, the target with the highest correlation with the target at the current time in different historical time periods is obtained as the correlation result; The final association matching result is obtained by piecing together the association results from all historical moments.
6. A multi-target tracking system based on a sparse attention mechanism, used to implement the multi-target tracking method based on a sparse attention mechanism as described in any one of claims 1 to 5, characterized in that, include: The target detection and target feature encoding module performs 3D target detection on the acquired multi-frame laser point cloud data, obtains the target detection results at the current time and the target detection results at the historical time, and encodes the target features at the current time and the target features at the historical time respectively. The sparse attention module constructs an attention matrix between the target features at the current time and the target features at historical time, calculates the attention mask between the target at the current time and the target at historical time, and obtains the sparse attention matrix between the target at the current time and the global target at all historical time through the attention mask; The state estimation and association matching output module calculates the state estimation result by weighting all target features at historical time points based on the sparse attention matrix. It then calculates the association results for all historical time points based on the sparse attention matrix at different times in the historical frame and uses these association results as the final association matching result.
7. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 5.
8. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 5.