Multi-target tracking method and device, electronic equipment and computer readable storage medium

Abstract

The invention provides a multi-target tracking method and device, electronic equipment and a computer readable storage medium, and the method comprises the steps: carrying out the target detection of a video sequence, and obtaining a detection frame of a current frame; under the condition that the detection frame of the current frame is successfully matched with the prediction frame, calculating a track consistency score; for the target trajectory of which the trajectory consistency score exceeds a set score, extracting feature representation of a historical frame by using a time sequence attention mechanism; and calculating feature similarity according to the feature representation and a set model, and determining the identity of the target trajectory. According to the method, the video frame is detected through the target detection technology, the corresponding high-confidence detection box is screened out, the purposes of accurately positioning multiple targets and reducing false detection can be achieved, and reliable input is provided for follow-up track association. The target position is predicted through the trajectory prediction technology, matching of a detection frame and a prediction frame is achieved, the purpose of maintaining trajectory continuity in a shielding or rapid motion scene is achieved, and the target identifier switching rate is remarkably reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of multi-target tracking, in particular, to a multi-target tracking method and device, an electronic device and a computer readable storage medium. BACKGROUND

[0002] Multi-target tracking, as one of the core tasks in the field of computer vision, has wide application value in intelligent monitoring, automatic driving, robot navigation and other fields. In recent years, with the breakthrough of deep learning technology, the tracking based on detection and joint detection tracking framework have significantly improved the tracking performance in the basic scene. However, in actual complex scenes, the existing methods still face two core bottlenecks: first, in the target dense, fast motion or severe occlusion scene, the tracking accuracy will decrease sharply, and the ID switching rate will usually increase by 3-5 times; second, the algorithm stability is sensitive to hyperparameters, and in the cross-camera, cross-resolution and other actual deployment scenes, the trajectory may be broken, and the frame rate stability of real-time tracking fluctuates more than 20%. These core bottlenecks greatly restrict the development and application of multi-target tracking. SUMMARY

[0003] Therefore, the embodiments of the present application aim to provide a multi-target tracking method and device, an electronic device and a computer readable storage medium, which can improve tracking stability while reducing target identification switching rate.

[0004] In a first aspect, the embodiments of the present application provide a multi-target tracking method, comprising: performing target detection on a video sequence to obtain a detection box of a current frame; in the case that the detection box of the current frame and a predicted box are matched successfully, calculating a trajectory consistency score; for a target trajectory whose trajectory consistency score exceeds a set score, using a time sequence attention mechanism to extract a feature representation of a historical frame; and calculating a feature similarity according to the feature representation and a set model to determine the identity of the target trajectory.

[0005] In the above implementation process, the target detection technology is used to detect the video frame, and the corresponding high-confidence detection box is selected, which can achieve the purpose of accurately positioning multiple targets and reducing false detection, and provide reliable input for subsequent trajectory association. The trajectory prediction technology is used to predict the target position, the matching of the detection box and the predicted box is realized, the purpose of maintaining the continuity of the trajectory in the occlusion or fast motion scene is achieved, and the target identification switching rate is significantly reduced.

[0006] In one embodiment, the trajectory consistency score is calculated in the case that the bounding box of the current frame matches the predicted bounding box, including: calculating the spatial position relationship between the bounding box of the current frame and the predicted bounding box; calculating the speed change amount of the historical speed of the target trajectory and the current speed of the target trajectory; calculating the cosine similarity of the direction vector of the historical speed of the target trajectory and the current speed of the target trajectory; calculating the trajectory consistency score based on the dynamic weight, the speed change amount and the cosine similarity.

[0007] In the above implementation process, through the trajectory consistency score, multiple indicators such as position continuity, speed smoothness and motion trend similarity are fused, the matching score is dynamically adjusted by weight, the purpose of comprehensively evaluating the rationality of target motion is achieved, and the trajectory association accuracy in complex scenes can be improved.

[0008] In one embodiment, for the target trajectory with a trajectory consistency score exceeding a set score, a time sequence attention mechanism is used to extract a feature representation of a historical frame, including: extracting appearance features of historical bounding boxes of the target trajectory to generate a feature sequence; modeling a time sequence relationship through a preset model to generate a context-enhanced feature representation; performing weight reduction processing on time frame features with a confidence lower than a set confidence through a confidence gate; and determining the feature representation according to the weighted processed feature sequence.

[0009] In the above implementation process, through the time sequence attention mechanism, the appearance feature sequence of the historical trajectory is modeled by a preset model to model the time sequence dependence, and the occluded frame features are dynamically reduced in weight through the confidence gate, so that the purpose of enhancing the robustness of the appearance features is achieved, the feature degradation problem caused by changes in light, blur or partial occlusion is effectively alleviated, and the accuracy of the appearance features is improved.

[0010] In one embodiment, the target detection is performed on the video sequence to obtain a bounding box of a current frame, including: performing multi-scale feature extraction on the video sequence to generate a candidate bounding box; eliminating overlapping boxes in the candidate bounding box through non-maximum suppression; and determining the bounding box remaining after the elimination of the overlapping boxes in the candidate bounding box as the bounding box of the current frame; wherein the bounding box of the current frame includes a bounding box position, a bounding box size and a bounding box category.

[0011] In the above implementation process, the video frame is detected through the target detection technology, and the high-confidence bounding box is screened out by combining the non-maximum suppression, so that multiple targets can be accurately positioned, the purpose of reducing false detection is achieved, and the reliability of the trajectory association is improved.

[0012] In one embodiment, the calculating the feature similarity according to the feature representation and the setting model, and determining the identity of the target track comprises: extracting a deep feature vector of the target track by a pedestrian re-identification model; calculating the similarity of the feature representation and the deep feature vector; and determining a track with a similarity higher than a similarity threshold as the same target.

[0013] In the above implementation process, by extracting the deep feature and calculating the similarity through the pedestrian re-identification technique, and determining the target identity in combination with the dynamic threshold, the appearance similar targets can be accurately distinguished, the cross-lens identity confusion can be reduced, the cross-scene tracking stability can be improved, and the application scenarios of the scheme are increased.

[0014] In one embodiment, the method further comprises: for an occluded track with an occlusion exceeding a set number of frames, predicting the target position by an optical flow method; supplementing the track information of the occluded track according to an interaction relationship between targets modeled by a graph neural network; and performing weight reduction processing on the attention weight of a low-quality frame.

[0015] In the above implementation process, by using the dynamic occlusion processing technology, the position of the track with long-time occlusion is predicted and supplemented by the optical flow, and the interaction relationship between targets is modeled by introducing the graph neural network, the lost track can be recovered in the group occlusion scene, and the accuracy of the target track identity recognition is improved.

[0016] In one embodiment, the method further comprises: processing the target detection model by channel pruning and / or int8 quantization; wherein the target detection model is configured to perform target detection on the video sequence; and / or processing the temporal attention module by block processing and / or sparse attention processing; wherein the temporal attention module is configured to extract the feature representation of the historical frame using the temporal attention mechanism.

[0017] In the above implementation process, by processing the target detection model by using the channel pruning and / or int8 quantization, the calculation amount can be reduced, and the multi-target tracking efficiency can be improved. In addition, by processing the temporal attention module by using the block processing and / or sparse attention, the calculation complexity of the long sequence can be reduced, and the calculation efficiency and accuracy can be improved.

[0018] In a second aspect, the embodiments of the present application also provide a multi-target tracking device, comprising: a detection module configured to perform target detection on a video sequence to obtain a detection box of a current frame; a calculation module configured to calculate a track consistency score in a case where the detection box of the current frame matches a predicted box successfully; an extraction module configured to extract a feature representation of a historical frame using a temporal attention mechanism for a target track with a track consistency score exceeding a set score; and a determination module configured to calculate a feature similarity according to the feature representation and a setting model, and determine the identity of the target track.

[0019] In a third aspect, the embodiments of the present application further provide an electronic device, comprising: a processor, a memory, wherein the memory stores machine readable instructions executable by the processor, and when the electronic device is running, the machine readable instructions are executed by the processor to perform the steps of the method in the first aspect or any possible implementation manner of the first aspect.

[0020] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and when the computer program is run by a processor, the steps of the multi-target tracking method in the first aspect or any possible implementation manner of the first aspect are performed.

[0021] In order to make the above objectives, characteristics and advantages of the present application more apparent, clear and easy to understand, the following will specifically describe embodiments, and make a detailed description below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments, and it should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation to the scope, and for those skilled in the art, other related drawings can also be obtained without paying creative labor on the basis of these drawings.

[0023] Figure 1 The block schematic diagram of the electronic device provided by the embodiments of the present application is shown in the figure; Figure 2 The flow chart of the multi-target tracking method provided by the embodiments of the present application is shown in the figure; Figure 3 The functional module schematic diagram of the multi-target tracking device provided by the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

[0025] It should be noted that: similar labels and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second" and the like are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0026] Multi-object tracking has wide applications in autonomous driving, intelligent surveillance, robot perception, etc. Currently, mainstream multi-object tracking methods mostly adopt the "detection-reassociation" paradigm, using deep neural networks to detect objects in images and associating these objects between different frames through appearance features or motion trajectories. However, when facing challenges such as occlusion, dense targets, fast motion, and similar appearance, the following problems exist: 1) High ID Switch rate: easy to lose or confuse identity when occlusion occurs.

[0027] 2) Insufficient utilization of historical trajectory: lack of effective modeling of temporal information.

[0028] 3) Poor robustness of appearance features: vulnerable to interference such as light and blur.

[0029] Therefore, the present application proposes a multi-object tracking method, which detects video frames through target detection technology, filters out corresponding high-confidence detection boxes, accurately locates multiple targets and reduces false positives, and provides reliable input for subsequent trajectory association. Through trajectory prediction technology, the target position is predicted, the matching of detection boxes and prediction boxes is realized, the continuity of trajectory in occlusion or fast motion scenarios is maintained, and the target identification switching rate is significantly reduced.

[0030] To facilitate the understanding of the present embodiment, first, the electronic device executing the multi-object tracking method disclosed in the present application is introduced in detail.

[0031] As shown in Figure 1 , it is a block diagram of an electronic device. The electronic device 100 can include a memory 111, a processor 113. Those skilled in the art can understand that Figure 1 the structure shown is only for illustration, and it does not limit the structure of the electronic device 100. For example, the electronic device 100 can include more or fewer components than those shown in Figure 1 , or have a different configuration than Figure 1 .

[0032] The above-mentioned memory 111, processor 113 are directly or indirectly electrically connected to each other to realize the transmission or interaction of data. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines. The above-mentioned processor 113 is used to execute the executable modules stored in the memory.

[0033] The memory 111 can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), and the like. The memory 111 is configured to store a program. After receiving an execution instruction, the processor 113 executes the program. The method performed by the electronic device 100 defined by the process disclosed in any of the embodiments of the present application can be applied to the processor 113 or implemented by the processor 113.

[0034] The processor 113 can be an integrated circuit chip having a signal processing capability. The processor 113 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), and the like. The processor 113 can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The processor 113 can implement or execute the methods, steps, and logical block diagrams disclosed in the embodiments of the present application. The general-purpose processor can be a microprocessor, or the processor can also be any conventional processor.

[0035] The electronic device 100 in the embodiments of the present application can be configured to execute each step in each method provided by the embodiments of the present application. The implementation process of the multi-target tracking method will be described in detail in the following several embodiments.

[0036] Please refer to Figure 2 , which is a flowchart of the multi-target tracking method provided by the embodiments of the present application. The specific process shown in FIG. 2 will be described in detail below. Figure 2

[0037] In step 201, target detection is performed on a video sequence to obtain a detection box of a current frame.

[0038] ​The video sequence refers to video information collected by an image collection device. The video sequence is video information input into an electronic device for executing the multi-target tracking method.

[0039] The target detection herein can be performed by setting a model. For example, a YOLOv8 model, a DETR model, etc. The model for target detection can be selected according to actual conditions. Alternatively, the target detection result of the video sequence can be through a detection box. For example, The detection box can include position coordinates and a confidence.

[0040] It should be understood that when the video sequence is subjected to target detection, target detection can be performed for each frame of the video sequence. In step 202, when the detection box of the current frame is successfully matched with the predicted box, the trajectory consistency score is calculated.

[0041] In an embodiment, before step 202, the method further includes performing position prediction on an existing trajectory to obtain a predicted box.

[0042] The position prediction on the existing trajectory can be achieved by a trajectory prediction model. For example, Kalman filtering, a long short-term memory network, etc. The trajectory prediction model can be selected according to actual conditions.

[0043] The Kalman filtering herein is an optimal estimation algorithm based on a linear system state equation, which is used for prediction of a uniform motion target and recursively predicts the target position through a state equation and an observation equation.

[0044] Exemplarily, the Kalman filtering parameters can include: state_transition=[1 0 1 0; 0 1 0 1; 0 0 1 0; 0 0 0 1]; % uniform model; process_noise=diag([0.1, 0.1, 0.5, 0.5]); The state_transition is a state transition matrix of a uniform motion model, which is commonly used for target tracking (such as vehicles, pedestrians, etc.). The process_noise is a process noise covariance matrix in Kalman filtering, which is used to reflect model uncertainty (such as unmodeled acceleration, measurement error, etc.).

[0045] The long short-term memory network is a special recurrent neural network specially designed to solve the gradient vanishing / explosion problem of traditional RNN when processing long sequence data, so as to effectively learn long-term dependencies. It is used for prediction of complex motion targets, trained through historical trajectory data to capture long-term motion dependencies.

[0046] The following is a specific implementation process of trajectory prediction by LSTM network: define a trajectory predictor class that inherits the PyTorch neural network module, create an LSTM layer with an input feature dimension of 4 (corresponding to x, y, vx, vy), a hidden layer dimension of 64, create a fully connected layer to map the 64-dimensional features output by LSTM to a 4-dimensional output. LSTM processes input sequences step by step, outputting the hidden state of all time steps. Take the output of the last time step and predict the state of the next time step through the fully connected layer. The above trajectory prediction model can be used to maintain an active trajectory list, recording the position, velocity and appearance features of the target.

[0047] Optionally, the detection frame and the prediction frame can be matched by Hungarian algorithm, KM algorithm, Hopcroft-Karp algorithm, auction algorithm, etc. The matching basis of the detection frame and the prediction frame can be intersection over union, Mahalanobis distance, etc. The matching method of the detection frame and the prediction frame can be selected according to actual conditions.

[0048] The trajectory consistency score here is an evaluation index for quantifying the spatio-temporal consistency between the predicted trajectory and the real trajectory, which comprehensively evaluates the matching degree of the trajectory from the overall trajectory shape, motion continuity, time synchronization, etc.

[0049] Step 203, for the target trajectory with trajectory consistency score exceeding the set score, use the time sequence attention mechanism to extract the feature representation of the historical frame.

[0050] Wherein, the set score can be set in advance according to the actual situation of the target trajectory. The target trajectory with trajectory consistency score exceeding the set score can be identified as a high-score trajectory.

[0051] The time sequence attention mechanism here is a key technology in deep learning for processing time series data, aiming to dynamically allocate weights to different time steps to capture long-term dependencies and key information in the sequence. The time sequence attention mechanism assigns weights to each time step in the sequence, dynamically adjusting the model's attention to different time points.

[0052] Wherein, the time sequence dependency can be modeled by Transformer or bidirectional gated recurrent unit, and robust feature representation can be generated.

[0053] In one embodiment, the feature representation of the historical frame can be extracted by the following way: using CNN to extract spatial features, then modeling time sequence relationship by Transformer, then dynamically aggregating key frame information, and combining memory mechanism to enhance long-term dependency, thereby realizing the extraction of the feature representation of the historical frame.

[0054] Step 204, calculating the feature similarity according to the feature representation and the setting model to determine the identity of the target trajectory.

[0055] The setting model can be a pedestrian re-identification model. The pedestrian re-identification model aims to determine whether it is the same target under different camera perspectives by analyzing the features in the image or video sequence.

[0056] It can be understood that the identity of the target trajectory can be determined as follows: the image or feature of the target trajectory in the video frame is identified by the pedestrian re-identification model, then the image or feature is mapped to a fixed-dimensional feature vector using a pre-trained RE target identification model, the feature vector is normalized to ensure scale consistency, and similarity calculation is performed. Finally, according to the similarity and the preset similarity threshold, the target trajectory with a similarity exceeding the similarity threshold is determined as the same target trajectory.

[0057] Optionally, the similarity calculation can be realized by cosine similarity, Euclidean distance, Mahalanobis distance, etc.

[0058] In the above implementation process, the target detection technology is used to detect the video frame, and the corresponding high-confidence detection box is selected, which can achieve the purpose of accurately positioning multiple targets and reducing false detection, and provide reliable input for subsequent trajectory association. The trajectory prediction technology is used to predict the target position, realize the matching of the detection box and the prediction box, achieve the purpose of maintaining the continuity of the trajectory in the occlusion or fast motion scene, and significantly reduce the target identification switching rate.

[0059] In one possible implementation, step 202 includes: calculating the spatial position relationship of the detection box and the prediction box of the current frame; calculating the speed change amount of the historical speed of the target trajectory and the current speed of the target trajectory; calculating the cosine similarity of the direction vector of the historical speed of the target trajectory and the current speed of the target trajectory; calculating the trajectory consistency score based on the dynamic weight, the speed change amount and the cosine similarity.

[0060] The spatial position relationship can be intersection over union, center point Euclidean distance, etc., which can be selected according to actual conditions.

[0061] In one embodiment, the calculation formula of the intersection over union can be: ; Wherein, is the coordinate region of the detection box, is the coordinate region of the prediction box, is the intersection over union.

[0062] The calculation formula of the center point Euclidean distance can be: ; wherein, is a center point coordinate of the detection box, is a center point coordinate of the prediction box, is a center point Euclidean distance.

[0063] The speed change amount of the historical speed and the current speed here can be determined by the speed difference between the historical speed and the current speed.

[0064] In an embodiment, a speed difference threshold can be set to constrain the speed difference.

[0065] Optionally, the cosine similarity can be calculated by the following formula: ; wherein, is a cosine similarity, is a current speed, is a historical speed.

[0066] The trajectory consistency score described above can be calculated by the following formula: ; wherein, is a dynamic weight, is a speed change amount, is a cosine similarity of the motion direction.

[0067] The dynamic weight here can be dynamically adjusted according to the actual situation, and then the scoring effect is optimized according to the adjustment of the dynamic weight.

[0068] Exemplarily, in the dense crowd monitoring scene, the strategy of dynamic adjustment of the trajectory consistency score weight can be: defining a function named adjust_weights, receiving a scene_type parameter (string type). If it is a crowded scene (crowd), the weight [0.4, 0.3, 0.3] is returned. If it is an occlusion scene (occlusion), the weight [0.3, 0.4, 0.3] is returned. Among them, the weights correspond to the evaluation dimensions of position continuity, speed smoothness and appearance similarity respectively.

[0069] In the above implementation process, by means of the trajectory consistency score, multiple indexes such as position continuity, speed smoothness and motion trend similarity are fused, the weight is dynamically adjusted to calculate the matching score, the purpose of comprehensively evaluating the rationality of the target motion is achieved, and the trajectory association accuracy in complex scenes can be improved.

[0070] In a possible implementation, step 203 comprises: extracting appearance features of the historical bounding boxes of the target trajectory to generate a feature sequence; modeling the time sequence relationship through a preset model to generate a context-enhanced feature representation; performing weight reduction processing on time frame features with a confidence lower than a set confidence through a confidence gate; and determining the feature representation according to the weighted processed feature sequence.

[0071] The appearance features can be extracted through a time sequence attention mechanism.

[0072] Understandably, the appearance features are extracted for each historical bounding box of the trajectory when the appearance features are extracted. The appearance features can also be arranged in a time sequence to generate a feature sequence. For example, the feature sequence can be: .

[0073] Optionally, the preset model can be a Transformer, a bidirectional gated recurrent unit, or the like, which can be selected according to actual conditions.

[0074] Understandably, the context-enhanced feature representation can be generated by using the internal mechanism (such as a cycle, attention, or convolution) of the preset model to capture the dependency (short-term / long-term, local / global) between sequence elements (time steps). When processing each sequence element, the preset model needs to consider not only the information of the element itself, but also the information of the context (previous and subsequent elements), and output a new feature vector (i.e., the context-enhanced feature representation) that integrates rich context.

[0075] It should be understood that the weight reduction processing on the time frame features with a confidence lower than a set confidence can reduce the influence of the feature representation with a low confidence on the result of the overall feature representation, and improve the accuracy of the feature representation.

[0076] In the above implementation process, the appearance feature sequence of the historical trajectory is modeled by using the preset model to model the time sequence dependency, and the occlusion frame features are dynamically reduced in weight through the confidence gate, which can achieve the purpose of enhancing the robustness of the appearance features, effectively alleviate the feature degradation problem caused by changes in light, blurring, or partial occlusion, and improve the accuracy of the appearance features.

[0077] In a possible implementation, step 201 comprises: performing multi-scale feature extraction on a video sequence to generate candidate bounding boxes; eliminating overlapping boxes in the candidate bounding boxes through non-maximum suppression; and determining the bounding boxes remaining after eliminating the overlapping boxes in the candidate bounding boxes as the bounding boxes of a current frame.

[0078] The bounding boxes of the current frame include a bounding box position, a bounding box size, and a bounding box category.

[0079] The multi-scale feature extraction step can be: extracting multi-layer feature maps through the network, upsampling the layer features, and splicing and fusing the low-layer features to form multi-scale features.

[0080] The non-maximum suppression described above is a technique of retaining the bounding box with the highest confidence and eliminating overlapping boxes.

[0081] In an embodiment, the non-maximum suppression can be implemented in the following way: the input detection boxes are sorted by confidence from high to low to obtain an index list after sorting. Initialize an empty list and save the final result. Select the box with the highest confidence, calculate the intersection over union of the box and the remaining boxes, remove the boxes with an intersection over union exceeding a threshold, update the index list of the remaining boxes, and repeat the above steps until all detection boxes are processed. Return the retained bounding box index or box coordinates.

[0082] In the implementation process described above, the video frame is detected by the target detection technology, and the high-confidence detection box is screened out by combining the non-maximum suppression, which can accurately locate multiple targets and reduce false positives, and improve the reliability of trajectory association.

[0083] In a possible implementation, step 204 includes: extracting a deep feature vector of the target trajectory by a pedestrian re-identification model; calculating the similarity of the feature representation and the deep feature vector; and determining that a trajectory with a similarity higher than a similarity threshold is the same target.

[0084] Optionally, the pedestrian re-identification model can be an ArcFace, OSNet or the like. The pedestrian re-identification model can be selected according to actual conditions.

[0085] The pedestrian re-identification model can detect the trajectory position frame by frame through a deep learning model, extract a global feature vector using a convolutional neural network, and then determine the deep feature vector of the target trajectory.

[0086] The similarity can be represented by cosine similarity, Euclidean distance, or the like. The determination method of the similarity can be selected according to actual conditions.

[0087] The similarity threshold is a similarity difference value that can be determined as the same trajectory. That is, the trajectories with a similarity within the similarity threshold range are the same trajectory.

[0088] Optionally, the similarity threshold can be dynamically adjusted according to actual conditions to adapt to the complexity of different scenes.

[0089] In the implementation process, the deep features are extracted and the similarity is calculated through the pedestrian re-identification technology, and the target identity is determined by combining the dynamic threshold, so as to accurately distinguish similar targets, reduce the identity confusion across the lens, improve the stability of cross-scene tracking, and increase the application scenarios of the scheme.

[0090] In a possible implementation, the method further includes: predicting the target position by an optical flow method for an occlusion trajectory whose occlusion exceeds a set number of frames; supplementing the trajectory information of the occlusion trajectory according to an interaction relationship between targets modeled by a graph neural network; and performing weight reduction processing on the attention weight of the low-quality frame.

[0091] The set number of frames can be set according to actual conditions. For example, the set number of frames can be 3 frames, 4 frames, 5 frames, etc.

[0092] In an embodiment, the prediction of the target position by the optical flow method can be implemented in the following manner: determining the target position by a target detection algorithm, calculating an optical flow field (which is a two-dimensional vector) between the current frame and the next frame, extracting the optical flow vector located in the target region from the calculated optical flow field, and aggregating the optical flow in the target region to obtain a vector representing the overall motion trend of the target. The position of the target in the next frame is predicted by the vector representing the overall motion trend.

[0093] The supplementing of the trajectory information of the occlusion trajectory according to the interaction relationship between targets modeled by the graph neural network can be implemented in the following manner: obtaining an original trajectory according to a target detection and tracking algorithm, marking the occlusion period in the original trajectory, and constructing a graph according to time slices and filling node and edge features. A pre-trained model is used to extract target appearance features and embed them into node features. The spatio-temporal graph is input, and the spatial GNN and the time sequence model are alternately optimized. For the occluded target, only the real position before and after the occlusion is used as a weak supervision signal. In the inference stage, the real position update of the occluded target is suspended, and the GNN output and the physical constraint are relied on to generate the trajectory.

[0094] It should be understood that the above-mentioned manner of supplementing the occlusion trajectory and predicting the target position is only exemplary, and the manner of supplementing the occlusion trajectory and predicting the target position can be adjusted accordingly according to actual conditions.

[0095] The above-mentioned weight reduction processing can be implemented by a time sequence attention mask. The time sequence attention mask suppresses the feature interference of the low-quality frame.

[0096] Exemplarily, the processing flow of the occlusion can be: initializing the network structure, calling the initialization method of the parent class nn. Module, defining a graph attention convolution layer, and supplementing the track information of the occluded track through the graph attention convolution layer. The parameters of the graph attention convolution layer can be as follows: the input feature dimension is 4 (for example, each node has a 4-dimensional feature), the output feature dimension is 8 (the feature dimension is expanded after the layer), and a multi-head attention mechanism containing 3 independent attention heads is used.

[0097] In the above implementation process, by adopting the dynamic occlusion processing technology, the position of the track with long-time occlusion is completed by optical flow prediction, and the interaction relationship between targets is modeled by introducing a graph neural network, so that the purpose of restoring the lost track in the group occlusion scene is achieved, and the accuracy of target track identity recognition is improved.

[0098] In a possible implementation, the method further includes: processing the target detection model through channel pruning and / or int8 quantization; and / or processing the temporal attention module through block processing and / or sparse attention.

[0099] The target detection model is configured to perform target detection on the video sequence, and the temporal attention module is configured to extract the feature representation of the historical frame using the temporal attention mechanism.

[0100] The channel pruning here is a structured model compression technology, which can significantly reduce the model calculation amount and parameter amount by removing redundant convolution channels and their corresponding feature maps in the neural network, while maintaining the integrity of the model structure to realize hardware-friendly acceleration.

[0101] Int8 quantization is a technology that converts high-precision floating-point parameters (such as FP32) in a deep learning model into 8-bit integers (INT8). It significantly reduces the memory occupancy and computational complexity of the model, and greatly improves the inference speed.

[0102] It should be understood that by adopting channel pruning and / or int8 quantization to process the target detection model, the calculation amount can be reduced, and the multi-target tracking efficiency can be improved.

[0103] The block processing described above is an optimization technology that divides long sequence data into multiple fixed-size blocks, processes each block separately to reduce computational complexity and memory occupancy, and aims to reduce computational complexity and reduce memory occupancy through local processing, while balancing computational efficiency and accuracy.

[0104] Sparse attention is a technology for optimizing traditional attention mechanisms, and its core idea is to limit the range of attention calculation and only focus on key parts of the input sequence, thereby reducing computational complexity and memory consumption.

[0105] It can be understood that by using the block processing and / or sparse attention processing timing attention module, the calculation complexity of the long sequence can be reduced, and the calculation efficiency and accuracy can be improved.

[0106] In an embodiment, the method further includes a plurality of optimization schemes, as shown in the following examples: 1. Fine-tune the RE target identification model through incremental training, so that the RE target identification model adapts to target appearance changes.

[0107] The core logic of fine-tuning the RE target identification model through incremental training is to realize the evolution of the RE target identification model through "staged unfreezing and local adjustment".

[0108] In an embodiment, the specific implementation steps can be: first, prepare incremental data, collect labeled data containing new target appearances (for the same target ID), and build a memory bank to store a small number of representative old samples (e.g., selected by clustering or hard example sampling, etc.). After loading the pre-trained RE target identification model, if the new data introduces a new target ID, expand the classification layer and initialize new neurons. In the training phase, mix the new data with the old samples in the memory bank in appropriate proportions, and use the knowledge distillation technique to constrain the model behavior: use the old model to generate soft labels (i.e., probability distribution) or feature vectors for the mixed data, design a composite loss function, and coordinate the learning intensity of new and old knowledge through weight coefficients. Fine-tune the entire model (or unfreeze the top layer parameters) with a small learning rate, and periodically evaluate the model's performance (e.g., mAP, Rank-1 indicators, etc.) on the test set containing historical appearances during the training process to monitor catastrophic forgetting. Update the model weights after training, and dynamically maintain the memory bank (e.g., replace redundant samples or add new category representatives, etc.), forming a closed loop of "data preparation → model expansion → mixed training → evaluation and update". Through the replay mechanism and distillation constraint, the model can absorb new appearance features while solidifying the memory of the original visual patterns, ultimately achieving stable re-identification capability in a long-term environment. 2. Reduce the feature distribution difference between simulated data and real scenes through adversarial domain adaptation technology.

[0109] The adversarial domain adaptation technology here is a transfer learning method that uses an adversarial training mechanism to reduce the distribution difference between different data domains. The core idea is to force the model to learn domain-invariant features through an adversarial game within the neural network, thereby transferring the knowledge learned in the labeled source domain (e.g., simulated data) to the unlabeled target domain (i.e., real scenes).

[0110] In an embodiment, the specific implementation steps can be: first, the labeled simulation data (source domain) and unlabeled real scene data (target domain) are prepared as input. If the real data is limited, the target domain samples can be expanded through data augmentation (such as random light change, adding noise). Then, the adversarial domain adaptation framework is selected: if feature-level alignment is required, DANN (Domain Adversarial Neural Network) can be used to make the feature extractor generate domain confusion features through the gradient reversal layer (GRL); if image-level style conversion is required, CycleGAN can be used to convert the simulation data into real scene style without paired data.

[0111] When building the model, DANN needs to share the feature extractor and add a domain classifier, and through adversarial training, the feature distribution is difficult to distinguish; CycleGAN needs two generators and discriminators, combined with adversarial loss and cycle consistency loss to ensure conversion quality. The training process is usually divided into pre-training and adversarial training stages: the pre-training stage trains the task model on the source domain, and the adversarial training stage aligns the feature distribution through GRL or GAN loss. In addition, DANN and CycleGAN can be used together to form an end-to-end framework to further improve the cross-domain performance.

[0112] In the evaluation stage, the target domain unlabeled data is used to verify the model performance (such as classification accuracy, detection mAP), and the t-SNE visualization feature space is used to check whether the source domain and target domain features overlap. 3, dynamically update the model parameters to improve the performance of cross-scene tracking.

[0113] Among them, the model parameters can be updated through online learning, or through transfer learning and domain adaptation, or through adaptive optimization and regularization. The dynamic updating method of the model parameters can be selected according to the actual situation.

[0114] Understandably, through dynamic updating of model parameters, in cross-scene tracking, the target may experience changes such as light changes, occlusions, deformations, or motion pattern mutations. For example, the scene conversion from indoor to outdoor will introduce changes in light intensity and background complexity, and dynamic updating of model parameters can incorporate new data features through online learning or incremental training, keeping the model adaptable to the current environment.

[0115] In addition, the data distribution of different scenes may have significant differences (such as resolution, viewing angle, noise level, etc.). Dynamic updating can gradually correct the bias caused by domain differences by adjusting the model parameters. For example, transfer learning combined with dynamic updating can reduce the data distribution difference between the source domain and the target domain, improving the generalization ability of cross-domain tracking.

[0116] 4. The trained deep learning model is compressed for precision (e.g., FP32 to FP16) to reduce the consumption of computing resources, improve inference speed, and maintain the model accuracy as much as possible.

[0117] Exemplarily, the precision compression of the trained deep learning model can be achieved through the following steps: first, use the command line tool provided by TensorRT for model conversion and performance testing. Then, specify the input ONNX format model file path and enable the half-precision floating point (FP16) mode to improve the calculation speed by about 3 times and reduce the memory occupation. Finally, save the optimized TensorRT engine as a binary file.

[0118] Understandably, by adopting the incremental training to fine-tune the RE target identification model and combining the adversarial domain adaptation to reduce the distribution difference, the purpose of self-adapting to target appearance changes and cross-scene migration can be achieved, and the cross-domain tracking accuracy is improved.

[0119] The above optimization scheme is only exemplary, and the optimization scheme of the multi-target tracking method can be adjusted according to actual conditions.

[0120] It should be understood that the multi-target tracking method in the embodiments of the present application integrates the detection, prediction, TCS, TAM and RE target identification modules through the multi-module cooperative technology to form a closed-loop system, achieves the purpose of optimizing the multi-target tracking performance end-to-end, and comprehensively improves the core indicators such as multi-target tracking accuracy. By deeply integrating the trajectory consistency and the time sequence attention mechanism, the purpose of achieving high-precision and high-robustness tracking in challenging scenarios such as occlusion, dense targets and fast motion is achieved, providing reliable technical support for applications such as autonomous driving and intelligent monitoring.

[0121] The following Table 1 shows the comparison of the delay of the target detection module and the time sequence attention mechanism module in the embodiments of the present application before and after optimization, to further demonstrate the effect brought by the technical scheme in the embodiments of the present application.

[0122] Table 1:

[0123] As can be seen from Table 1, whether the target detection is optimized or the time sequence attention mechanism is optimized, the delay after optimization can be reduced, and the processing efficiency can be improved.

[0124] In the above implementation process, by adopting channel pruning and / or int8 quantization to process the target detection model, the calculation amount can be reduced, and the multi-target tracking efficiency can be improved. In addition, by adopting block processing and / or sparse attention processing of the time sequence attention module, the calculation complexity of long sequences can be reduced, and the calculation efficiency and accuracy can be improved.

[0125] Based on the same application concept, the present embodiment also provides a multi-target tracking device corresponding to the multi-target tracking method. Since the device in the present embodiment solves the problem by a similar principle to the multi-target tracking method in the foregoing embodiment, the implementation of the device in the present embodiment can refer to the description in the foregoing method embodiment, and the repeated parts will not be described herein.

[0126] Please refer to Figure 3 FIG. 1 is a functional module schematic diagram of a multi-target tracking device provided in the present embodiment. Each module in the multi-target tracking device in the present embodiment is configured to perform each step in the foregoing method embodiment. The multi-target tracking device includes a detection module 301, a calculation module 302, an extraction module 303, and a determination module 304. The detection module 301 is configured to perform target detection on a video sequence to obtain a detection box of a current frame.

[0127] The calculation module 302 is configured to calculate a trajectory consistency score in a case where the detection box of the current frame and a predicted box are successfully matched.

[0128] The extraction module 303 is configured to use a time sequence attention mechanism to extract a feature representation of a historical frame for a target trajectory whose trajectory consistency score exceeds a set score.

[0129] The determination module 304 is configured to calculate a feature similarity according to the feature representation and a set model, and determine an identity of the target trajectory.

[0130] In one possible implementation, the calculation module 302 is further configured to: calculate a spatial position relationship between the detection box of the current frame and the predicted box; calculate a speed change amount of a historical speed of the target trajectory and a current speed of the target trajectory; calculate a cosine similarity of a direction vector of the historical speed of the target trajectory and the current speed of the target trajectory; and calculate the trajectory consistency score based on a dynamic weight, the speed change amount, and the cosine similarity.

[0131] In one possible implementation, the extraction module 303 is further configured to: extract an appearance feature of a historical detection box of the target trajectory to generate a feature sequence; model a time sequence relationship through a preset model to generate a context-enhanced feature representation; perform weight reduction processing on a time frame feature whose confidence is lower than a set confidence through confidence gating; and determine the feature representation according to the weighted processed feature sequence.

[0132] In a possible implementation, the detection module 301 is further configured to perform multi-scale feature extraction on the video sequence to generate candidate detection boxes, eliminate overlapping boxes in the candidate detection boxes through non-maximum suppression, and determine the detection boxes remaining after the elimination of the overlapping boxes as detection boxes of the current frame, wherein the detection boxes of the current frame include detection box positions, detection box sizes, and detection box categories.

[0133] In a possible implementation, the determination module 304 is further configured to extract a deep feature vector of the target trajectory through a pedestrian re-identification model, calculate the similarity between the feature representation and the deep feature vector, and determine that a trajectory with a similarity higher than a similarity threshold is the same target.

[0134] In a possible implementation, the multi-target tracking apparatus further includes a first processing module configured to predict target positions through an optical flow method for occluded trajectories with an occlusion number exceeding a set number of frames, supplement trajectory information of the occluded trajectories according to an interaction relationship between targets modeled by a graph neural network, and perform weight reduction processing on attention weights of low-quality frames.

[0135] In a possible implementation, the multi-target tracking apparatus further includes a second processing module configured to process a target detection model through channel pruning and / or int8 quantization, wherein the target detection model is configured to perform target detection on a video sequence, and / or process a temporal attention module through block processing and / or sparse attention processing, wherein the temporal attention module is configured to extract feature representations of historical frames using a temporal attention mechanism.

[0136] In addition, the embodiment of the present application further provides a computer readable storage medium, and the computer readable storage medium stores a computer program. When the computer program is run by a processor, the steps of the multi-target tracking method described in the above method embodiments are executed.

[0137] The computer program product of the multi-target tracking method provided by the embodiment of the present application includes a computer readable storage medium storing program codes. The program codes include instructions for executing the steps of the multi-target tracking method described in the above method embodiments. For details, refer to the above method embodiments, which will not be described here again.

[0138] It should be understood that all the functional units in the embodiments of the present application can be integrated into one independent part, or each functional unit can exist alone, or two or more functional units can be integrated into one independent part.

[0139] In addition, each functional module in the various embodiments of the present application can be integrated together to form an independent part, or each module can exist alone, or two or more modules can be integrated to form an independent part.

[0140] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or parts 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.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: 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, and various media that can store program codes. It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the elements defined by the statement "include" do not exclude the presence of other identical elements in the processes, methods, articles or devices that include the elements. The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0141] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

Claims

1. A multi-target tracking method, characterized by, The method comprises: target detection is performed on a video sequence to obtain a detection box of a current frame; in a case where the detection box of the current frame matches a predicted box successfully, a trajectory consistency score is calculated; for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory.

2. The method of claim 1, wherein, The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; 3. The method according to claim 1 or 2, characterized in that, a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises:

4. The method of claim 1, wherein, for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; 5. The method of claim 1, wherein, a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory.

6. The method according to any one of claims 1 to 5, characterized in that, The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises:

7. The method according to any one of claims 1 to 5, characterized in that, for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises:

8. A multi-target tracking device, characterized by, for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency score exceeds a set score, a feature representation of a historical frame is extracted using a time sequence attention mechanism; a feature similarity is calculated according to the feature representation and a set model to determine the identity of the target trajectory. The method further comprises: for a target trajectory whose trajectory consistency An extraction module is configured to extract a feature representation of a historical frame using a temporal attention mechanism for a target trajectory whose trajectory consistency score exceeds a set score; A determination module is configured to calculate a feature similarity according to the feature representation and a set model, and determine an identity of the target trajectory.

9. An electronic device, comprising: The method comprises the steps of: A processor and a memory, wherein the memory stores machine readable instructions executable by the processor, and when the electronic device is running, the machine readable instructions are executed by the processor to perform the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer readable storage medium stores a computer program, and when the computer program is run by a processor, the steps of the method according to any one of claims 1 to 7 are performed.

Images