Multi-target tracking method of target domain data based on spatial topology and identity aggregation

Abstract

This invention provides a method for the evolution of a multi-target tracking network based on spatial topology and identity aggregation. The method includes: (1) training a multi-target tracking network and an identity aggregation network based on spatial topology in a labeled source domain to obtain parameter-updated multi-target tracking and identity aggregation networks; (2) using the parameter-updated multi-target tracking and identity aggregation networks to obtain pseudo-labels for data in an unlabeled target domain; and (3) using the obtained pseudo-labels and the self-supervised module in the multi-target tracking network to update the parameters of the entire multi-target tracking network. The network parameters obtained after iteratively executing steps (2) and (3) for multiple rounds are used for multi-target tracking in the target domain. This invention can solve the domain adaptation problem in multi-target tracking and suppress the significant performance degradation of multi-target tracking network models in unlabeled tracking scenarios.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a multi-target tracking method based on target domain data of spatial topology and identity aggregation. Background Technology

[0002] With the development of deep learning, convolutional neural networks are being applied in an increasing number of scenarios. Multi-object tracking, due to its widespread use in video surveillance, human-computer interaction, and virtual reality, is receiving increasing attention in the field of computer vision. Multi-object tracking aims to locate multiple target objects in a given video sequence, assigning different IDs to different objects and recording the trajectory of each ID in the video. Currently, with the continuous development of object detection technology based on convolutional neural networks, detection-based tracking algorithms have become the mainstream direction for multi-object tracking. Detection-based tracking algorithms first need to perform object detection on each video frame to obtain the detection results for each frame, and then perform data association based on the detection results to create the trajectory of each object in the video.

[0003] Traditional two-stage multi-object tracking algorithms treat detection and tracking as two independent models. First, a detection model is used to obtain detection results, and then these results, along with the original video, are input into the tracking model for multi-object tracking. Since the detection and tracking models are executed separately, this leads to resource waste and can potentially cause efficiency issues. In recent years, with the rapid maturation of deep learning technology, one-stage multi-object tracking frameworks have attracted more research attention. Their core idea is to consider detection and tracking simultaneously within a single network. These methods have brought significant improvements in efficiency and accuracy for multi-object tracking tasks. However, due to the high cost of annotation, one-stage multi-object tracking methods remain difficult to apply in real-world scenarios. Therefore, how to design a one-stage multi-object tracking model that can adapt to new label-free tracking scenarios remains an unexplored and challenging problem that needs to be solved.

[0004] One existing one-stage multi-target tracking method involves simultaneously performing target detection and tracking within a deep learning network, reducing inference time by sharing most of the computational resources. The drawback of this method is that it primarily focuses on tracking efficiency while neglecting the requirement for tracking accuracy.

[0005] Another existing one-stage multi-target tracking method that simultaneously balances tracking efficiency and accuracy uses a ResNet-34 structure as a feature extractor and adds a DLA (Deep Layer Aggregation) structure to the feature extractor to achieve detection at different scales. This network design can dynamically adapt the receptive field according to the target's scale and pose, while also helping to alleviate alignment problems.

[0006] Based on the feature extractor, this method adds two branches: object detection and identity embedding, to achieve one-stage multi-object tracking. The object detection branch is viewed as a center-based bounding box regression task on a high-resolution feature map. Three parallel regression heads are added to the feature extractor to predict the heatmap, target bounding box size, and center offset, respectively. Each head is implemented by performing a 3×3 convolution (256 channels) on the output feature map of the feature extractor, followed by a 1×1 convolutional layer to generate the final target. The identity embedding branch aims to generate features that can distinguish different targets. Ideally, the distance between different targets should be greater than the distance between the same target. To achieve this, a convolutional layer with 128 kernels is applied to the feature map obtained by the feature extractor to extract identity embedding features at each location. Finally, this method achieves multi-object tracking based on the network output detection values ​​and identity embedding features. The target trajectory is initialized based on the detection boxes in the first frame. In subsequent frames, the detection values ​​are associated with the existing trajectories based on the identity embedding features and the trajectory distance measured by IoU to achieve multi-object tracking. The drawbacks of this method include: its performance benefits from a large amount of training data; however, in practical applications, it is impractical to obtain a large amount of time-consuming and labor-intensive training data; and if the trained model is directly applied to a new tracking scenario, its performance will drop significantly. Summary of the Invention

[0007] The embodiments of the present invention provide a multi-target tracking network evolution method based on spatial topology and identity aggregation to solve the domain adaptation problem of multi-target tracking and improve the tracking performance of multi-target tracking network models in unlabeled target domains.

[0008] To achieve the above objectives, the present invention adopts the following technical solution.

[0009] A method for multi-target tracking network evolution based on spatial topology and identity aggregation includes:

[0010] Step S1: Train a multi-target tracking network and an identity aggregation network based on spatial topology in a labeled source domain to obtain the multi-target tracking network and identity aggregation network with updated parameters.

[0011] Step S2: Use the updated multi-target tracking network and identity aggregation network to obtain pseudo-labels for unlabeled target domain data;

[0012] Step S3: Update the parameters of the multi-target tracking network using the pseudo-labels of the unlabeled target domain data and the self-supervised module in the multi-target tracking network.

[0013] Preferably, training a multi-target tracking network and an identity aggregation network based on spatial topology in a labeled source domain to obtain a parameter-updated multi-target tracking network and an identity aggregation network includes:

[0014] A one-stage multi-target tracking model framework is constructed, which includes a multi-target tracking network module based on spatial topology and an identity aggregation network module. The multi-target tracking network includes a feature extractor, a detection branch, an identity embedding branch, and a self-supervised context prediction branch. The identity aggregation network learns the weights of target features in different data frames and fuses the target features of historical data frames according to the weights, and uses the fused target features as the feature representation of the target.

[0015] Labeled source domain sequence data is input into a multi-object tracking network based on spatial topology. The multi-object tracking network uses a feature extractor to extract global features from the labeled source domain sequence data, calculates the loss functions of the detection branch, identity embedding branch, and self-supervised context prediction branch, and adjusts the network parameters based on backpropagation according to the loss functions to obtain the multi-object tracking network with updated parameters trained in the source domain.

[0016] The encoded source domain sequence data is extracted using the trained multi-target tracking network. The encoded data is then input into the identity aggregation network. The identity aggregation network aggregates historical embeddings with the same identity by using different learning weights. The parameters of the entire identity aggregation network are adjusted by backpropagation according to the loss function of the identity aggregation network, thus obtaining the updated identity aggregation network with parameters trained in the source domain.

[0017] Preferably, the method of obtaining pseudo-labels for unlabeled target domain data using a multi-target tracking network and an identity aggregation network with updated parameters includes:

[0018] Each frame of data in the unlabeled target domain is sequentially input into the multi-target tracking network and the identity aggregation network after parameter updates. The multi-target tracking network uses a feature extractor to extract global features from the data frame, uses a detection branch to detect targets of interest in the data frame, and uses an identity embedding branch to obtain discriminative features in the data frame that can distinguish different identity objects. The targets obtained by the detection branch in the data frame constitute the detection value of each frame of data, and the discriminative features obtained by the identity embedding branch constitute the features of the detection value. The identity aggregation network automatically learns the weights of historical embeddings and aggregates embeddings with the same label according to the learned weights. Low attention weights are assigned to blurry, incorrect, or partially occluded detections to obtain the target feature representation of each frame of data.

[0019] A bipartite graph is generated based on the features of the detected values ​​and the target features of each frame of data. The Hungarian algorithm is used to match the bipartite graph to achieve data association between the detected values ​​and the tracked targets. The associated detected values ​​are added to the corresponding targets to update the tracked targets. Finally, all the targets constitute pseudo-labels in the target domain.

[0020] Preferably, the step of updating the parameters of the multi-target tracking network using pseudo-labels from the unlabeled target domain data and the self-supervised module in the multi-target tracking network includes:

[0021] The context features of unlabeled target domain data are learned by utilizing the self-supervised context prediction branch in the multi-target tracking network. The parameters of the feature extractor, detection branch and identity embedding branch in the multi-target tracking network are updated by using the pseudo-labels of the unlabeled target domain data and the self-supervised context prediction branch.

[0022] The above steps S2 and S3 are executed repeatedly to continuously update the parameters of the multi-target tracking network, so that the updated feature extractor, detection branch and identity embedding branch are adapted to the data distribution of the target domain.

[0023] As can be seen from the technical solutions provided by the embodiments of the present invention above, the embodiments of the present invention design a multi-target tracking network evolution method based on spatial topology and identity aggregation, which can solve the domain adaptation problem of one-stage multi-target tracking and suppress the problem of significant performance degradation of multi-target tracking network models in unlabeled tracking scenarios.

[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A schematic diagram of the structure of a one-stage multi-target tracking network evolution model framework based on spatial topology and identity aggregation provided in an embodiment of the present invention;

[0027] Figure 2 This is a flowchart illustrating the processing of a multi-target tracking network evolution method based on spatial topology and identity aggregation, as provided in an embodiment of the present invention. Detailed Implementation

[0028] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0029] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0030] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0031] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0032] This invention proposes a multi-target tracking network evolution method based on spatial topology and identity aggregation to improve the generalization ability of multi-target tracking networks in the target domain (i.e., tracking scenarios without data labels for network training). Specifically, a one-stage tracking network based on spatial topology is designed to learn the spatial topological relationships of the tracking scenario through a self-supervised mechanism, thereby obtaining discriminative feature representations without relying on labeled data. An identity aggregation network is also designed to fuse features of each target in different data frames to provide a more reliable target representation. In the target domain, the multi-target tracking network based on spatial topology and the identity aggregation network iteratively perform pseudo-label collection and parameter updates to achieve network evolution from the source domain to the target domain. The evolved network can capture the distribution of the target domain, thus avoiding a significant performance degradation caused by the different distributions of the source and target domains.

[0033] Before development, engineers need to configure the experimental environment, including installing the Linux operating system, PyCharm software, and the PyTorch deep learning framework. Furthermore, because deep learning model training is time-consuming, it's best to train on a GPU. During development, careful attention must be paid to the code implementation of the forward propagation formula corresponding to the loss function. The most challenging aspect of the entire development process is controlling the settings of hyperparameters, such as thresholds for different modules, learning rates, and the number of iterations, as these parameters significantly impact the experimental results.

[0034] Compared to two-stage multi-object tracking, which separates object detection and tracking into two deep learning networks, the core idea of ​​one-stage multi-object tracking is to perform object detection and tracking simultaneously in a single deep learning network, thereby reducing inference time by sharing most of the computational resources.

[0035] The structural diagram of the one-stage multi-target tracking model framework provided in this embodiment of the invention is shown below. Figure 1 As shown, the system mainly includes a multi-target tracking network module based on spatial topology and an identity aggregation network module. The multi-target tracking network based on spatial topology further includes a feature extractor, a detection branch, an identity embedding branch, and a context prediction branch. This invention inputs labeled source domain sequences and unlabeled target domain sequences into the aforementioned one-stage multi-target tracking model framework. Through pseudo-label collection and parameter updates, the network evolves, thereby achieving domain adaptation for multi-target tracking.

[0036] A multi-target tracking network based on spatial topology is used to learn the potential distribution from the source domain to the target domain. It includes a feature extractor, a detection branch, an identity embedding branch, and a self-supervised context prediction branch, which learns the context features of the unlabeled inference domain while generating pseudo-labels.

[0037] The feature extractor is responsible for extracting global features from the data frame. Based on the feature extractor, a detection branch and an identity embedding branch are constructed to complete a multi-object tracking operation. The detection branch detects the target of interest, while the identity embedding branch obtains discriminative features that distinguish different identities. The self-supervised context prediction branch is based on spatial topology; it can associate spatial context by predicting the relative positions of the central block and its neighbors without labeled data, thereby helping the feature extractor learn discriminative features.

[0038] The identity aggregation network automatically learns the weights of historical embeddings and aggregates embeddings with the same identifier based on the learned weights. It assigns low attention weights to blurry, incorrect, or partially occluded detections, thereby reducing the impact of low-quality object bounding boxes during tracking and suppressing noisy labels during network evolution.

[0039] This invention uses the deep learning framework PyTorch to build a network framework, and implements the evolution of the target domain network for multi-target tracking tasks in three steps:

[0040] (1) Train a multi-target tracking network and an identity aggregation network based on spatial topology in a labeled source domain;

[0041] (2) Use the parameter-updated multi-target tracking network and identity aggregation network to obtain pseudo-labels for data in the unlabeled target domain;

[0042] (3) Update the parameters of the entire multi-target tracking network using the obtained pseudo-labels and the self-supervised module in the multi-target tracking network.

[0043] The network parameters obtained after iteratively executing steps (2) and (3) above for multiple rounds are used for multi-target tracking in the target domain.

[0044] based on Figure 1 The model framework shown in this invention, and the processing flow of the multi-target tracking network evolution method based on spatial topology and identity aggregation provided in this embodiment are as follows: Figure 2 As shown, the processing steps include the following:

[0045] Step S10. Train a multi-target tracking network and an identity aggregation network based on spatial topology in the labeled source domain.

[0046] Since the one-stage multi-object tracking network trained on the labeled domain needs to be applied to the unlabeled domain, the model must first be trained in the labeled source domain, that is, the multi-object tracking network and identity aggregation network based on spatial topology are trained in the labeled source domain.

[0047] Labeled source domain sequences are input into a spatial topology-based multi-object tracking network. The multi-object tracking network uses a feature extractor to extract global features, calculates loss functions for the detection branch, identity embedding branch, and self-supervised context prediction branch, and adjusts network parameters through backpropagation based on the loss functions, resulting in a multi-object tracking network trained in the source domain. The trained multi-object tracking network is then used to extract target encodings, which are input into an identity aggregation network. The parameters of the entire identity aggregation network are adjusted through backpropagation based on the aggregation network's loss function.

[0048] Since the poses of some objects change constantly in multi-object tracking tasks, and they may even be occluded by other objects, this invention implements an identity aggregation network. This identity aggregation network can aggregate historical embeddings with the same identity by using different learning weights, providing a more reliable object representation for multi-object tracking networks based on spatial topology.

[0049] Step S20. Use the parameter-updated multi-target tracking network and identity aggregation network to obtain pseudo-labels for the data in the unlabeled target domain.

[0050] Because the source and target domains have different distributions, networks trained in the source domain may experience a significant performance drop when tracking objects in the target domain. Therefore, to track objects in the unlabeled target domain, sequence frames from the unlabeled target domain need to be sequentially input into the updated model. This involves continuously collecting pseudo-labels and updating parameters to evolve the network, enabling it to capture the distribution of the target domain and thus avoiding the significant performance degradation caused by the different distributions of the source and target domains. The pseudo-labels are generated using a feature extractor, a detection branch, and an identity embedding branch.

[0051] Each frame of data from the unlabeled target domain is sequentially input into a multi-object tracking network (using only a feature extractor, detection branch, and identity embedding branch) and an identity aggregation network after parameter updates. The multi-object tracking network uses the feature extractor to extract global features from the data frame, the detection branch to detect targets of interest in the data frame, and the identity embedding branch to obtain discriminative features that can distinguish different identities in the data frame. The targets obtained by the detection branch constitute the detection values ​​of each frame, and the discriminative features obtained by the identity embedding branch constitute the features of the detection values.

[0052] The identity aggregation network automatically learns the weights of historical embeddings and aggregates embeddings with the same identifier based on the learned weights. It assigns low attention weights to detections that are blurry, incorrect, or partially occluded, thus obtaining the target feature representation of each frame of data.

[0053] A bipartite graph is generated based on the features of the detected values ​​and the target features of each frame of data. The Hungarian algorithm is used to match the bipartite graph to realize the data association between the detected values ​​and the tracked targets. The associated detected values ​​are added to the corresponding targets to update the tracked targets. Finally, all the targets constitute pseudo-labels in the target domain.

[0054] Step S30. Update the parameters of the entire multi-object tracking network using the obtained pseudo-labels and the self-supervised module in the multi-object tracking network.

[0055] The obtained pseudo-labels and self-supervised context prediction branch are used to update the parameters of the feature extractor, detection branch, and identity embedding branch in the multi-target tracking network, so that the updated feature extractor, detection branch, and identity embedding branch are adapted to the data distribution of the target domain, thereby improving the tracking performance of the unlabeled target domain.

[0056] The steps S20 and S30 are executed repeatedly to continuously update the parameters and improve the tracking performance of the multi-target tracking network in the unlabeled target domain.

[0057] This invention incorporates a self-supervised mechanism to learn spatial context features and proposes an identity aggregation network. Because the target domain lacks labeled data, directly applying a model trained in a labeled source domain to an unlabeled target domain results in a significant drop in performance. To improve the generalization ability of the multi-object tracking model in the unlabeled target domain, this invention adds a self-supervised mechanism to learn spatial context features. With the help of this self-supervised branch, the multi-object tracker in this invention can associate spatial context by predicting the relative positions between the central region and its neighboring regions. This helps learn discriminative features and capture the spatial topological information of the object.

[0058] In summary, this invention presents a method for multi-target tracking network evolution based on spatial topology and identity aggregation. The multi-target tracking network model based on spatial topology can adapt to new label-free tracking scenarios through network evolution, and the identity aggregation network can aggregate historical embeddings with the same identity by using different learning weights, thus providing a more reliable object representation for the multi-target tracking network based on spatial topology.

[0059] In one-stage multi-target tracking based on deep learning, a large amount of labeled data is required to train a high-performance tracking model. However, obtaining a large amount of time-consuming and labor-intensive training data is impractical in real-world applications. Moreover, due to the different distributions in different tracking scenarios, applying a multi-target tracking network model trained in one tracking scenario to a completely new tracking scenario often results in a significant performance drop. Therefore, this invention proposes a multi-target tracking network evolution method based on spatial topology and identity aggregation. This method utilizes a model trained in one tracking scenario (source domain) and unlabeled data from a completely new tracking scenario (target domain) to evolve the network. This allows the evolved network to capture the distribution of the target domain, thereby improving tracking performance in entirely new tracking scenarios.

[0060] This invention designs a multi-target tracking network evolution method based on spatial topology and identity aggregation. The spatial topology-based multi-target tracking network model can adapt to new label-free tracking scenarios through network evolution. This invention implements an identity aggregation network that can aggregate historical embeddings with the same identity using different learning weights, providing a more reliable object representation for spatial topology-based multi-target tracking networks.

[0061] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0062] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0063] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0064] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for multi-target tracking network evolution based on spatial topology and identity aggregation, characterized in that, include: Step S1: Train a multi-target tracking network and an identity aggregation network based on spatial topology in a labeled source domain to obtain the multi-target tracking network and identity aggregation network with updated parameters. Step S2: Use the updated multi-target tracking network and identity aggregation network to obtain pseudo-labels for the unlabeled target domain data; Step S3: Update the parameters of the multi-target tracking network using the pseudo-labels of the unlabeled target domain data and the self-supervised module in the multi-target tracking network; The process of training a spatial topology-based multi-target tracking network and an identity aggregation network in a labeled source domain to obtain parameter-updated multi-target tracking and identity aggregation networks includes: A one-stage multi-target tracking model framework is constructed, which includes a multi-target tracking network module based on spatial topology and an identity aggregation network module. The multi-target tracking network includes a feature extractor, a detection branch, an identity embedding branch, and a self-supervised context prediction branch. The identity aggregation network learns the weights of target features in different data frames and fuses the target features of historical data frames according to the weights, and uses the fused target features as the feature representation of the target. Labeled source domain sequence data is input into a multi-object tracking network based on spatial topology. The multi-object tracking network uses a feature extractor to extract global features from the labeled source domain sequence data, calculates the loss functions of the detection branch, identity embedding branch, and self-supervised context prediction branch, and adjusts the network parameters based on backpropagation according to the loss functions to obtain the multi-object tracking network with updated parameters trained in the source domain. The encoded source domain sequence data is extracted using the trained multi-target tracking network. The encoded data is then input into the identity aggregation network. The identity aggregation network aggregates historical embeddings with the same identity by using different learning weights. The parameters of the entire identity aggregation network are adjusted by backpropagation according to the loss function of the identity aggregation network, and the updated identity aggregation network obtained from the parameters trained in the source domain is obtained. The method of obtaining pseudo-labels for unlabeled target domain data using a parameter-updated multi-target tracking network and an identity aggregation network includes: Each frame of data in the unlabeled target domain is sequentially input into the multi-target tracking network and the identity aggregation network after parameter updates. The multi-target tracking network uses a feature extractor to extract global features from the data frame, uses a detection branch to detect targets of interest in the data frame, and uses an identity embedding branch to obtain discriminative features in the data frame that can distinguish different identity objects. The targets obtained by the detection branch in the data frame constitute the detection value of each frame of data, and the discriminative features obtained by the identity embedding branch constitute the features of the detection value. The identity aggregation network automatically learns the weights of historical embeddings and aggregates embeddings with the same label according to the learned weights. Low attention weights are assigned to blurry, incorrect, or partially occluded detections to obtain the target feature representation of each frame of data. A bipartite graph is generated based on the features of the detected values ​​and the target features of each frame of data. The Hungarian algorithm is used to match the bipartite graph to achieve data association between the detected values ​​and the tracked targets. The associated detected values ​​are added to the corresponding targets to update the tracked targets. Finally, all the targets constitute pseudo-labels in the target domain.

2. The method according to claim 1, characterized in that, The method of using pseudo-labels from the unlabeled target domain data and a self-supervised module in the multi-target tracking network to update the parameters of the multi-target tracking network includes: The context features of unlabeled target domain data are learned by utilizing the self-supervised context prediction branch in the multi-target tracking network. The parameters of the feature extractor, detection branch and identity embedding branch in the multi-target tracking network are updated by using the pseudo-labels of the unlabeled target domain data and the self-supervised context prediction branch. The above steps S2 and S3 are executed repeatedly to continuously update the parameters of the multi-target tracking network, so that the updated feature extractor, detection branch and identity embedding branch are adapted to the data distribution of the target domain.

Images