A nighttime multi-target tracking method and system based on domain adaptive learning

Abstract

The present invention belongs to the field of computer vision technology and provides a method and system for nighttime multi-target tracking based on domain adaptive learning, comprising: acquiring daytime image data and nighttime image data; performing image conversion processing on the acquired daytime image data to obtain a nighttime converted image; performing a nighttime multi-target tracking task based on the nighttime image data, the nighttime converted image, and a domain adaptive learning model to complete multi-target tracking in a nighttime environment; wherein the domain adaptive learning model adopts a teacher-student network, obtains tracking results based on the teacher network and the nighttime image data, and uses the obtained tracking results as pseudo labels; based on the pseudo labels, the student network is guided to perform unsupervised learning on the nighttime image data, the student network performs supervised learning on the nighttime converted image, and the student network is trained according to a multi-domain learning loss value of the student network to obtain a domain adaptive learning model.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a nighttime multi-target tracking method and system based on domain adaptive learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of deep learning, neural network technology is being applied in an increasing number of scenarios. Multi-target tracking, due to its widespread application in video surveillance, autonomous driving, and intelligent transportation, is attracting increasing attention in the field of computer vision. Its primary goal is to accurately track multiple moving targets in a video sequence while maintaining their identity consistency. With the continuous advancement of convolutional neural network target detection technology, detection-based tracking algorithms have become the mainstream in the field of multi-target tracking. These algorithms use target detection to identify targets in each frame, then use these detection results to perform data association to establish the target's trajectory in the video sequence.

[0004] Existing multi-target tracking methods are mainly developed for daytime data, which results in relatively poor multi-target tracking performance in nighttime environments. Due to the difficulty and high cost of data collection in nighttime environments and the complexity of labeling, labeled nighttime datasets are very scarce. The lack of sufficient quantity and quality of labeled nighttime datasets limits the performance improvement of multi-target tracking algorithms in nighttime environments. There are significant differences between daytime and nighttime environments. Changes in factors such as illumination, shadows, and reflections lead to large differences in brightness and contrast. Due to the scarcity of labeled nighttime data, supervised learning methods are difficult to implement and are prone to overfitting, which makes it difficult to improve the performance of the algorithm. Therefore, how to deal with nighttime datasets and occluded targets has become a major problem that needs to be solved urgently.

[0005] Existing multi-target tracking relies on supervised learning with large amounts of labeled data, but performance often degrades significantly in new, unlabeled data scenarios. A multi-target tracking method that uses unsupervised domain adaptive learning significantly improves performance in new scenarios by leveraging pseudo-labels. This method constructs a multi-target tracking network using a feature extractor, a detection branch, an identity embedding branch, and a context prediction branch. This network is used to obtain pseudo-labels for data in new scenarios, which are then corrected using an identity aggregation module. The entire multi-target tracking network is then trained using the pseudo-labels combined with the context prediction branch. The resulting parameter-updated network is then used to track multiple targets in new scenarios. This method effectively improves tracking performance in unlabeled scenarios, but it is primarily applied to daytime scenes, such as campuses, squares, and roads, ignoring the effects of illumination variations between different scenes. Consequently, performance remains suboptimal in nighttime scenes with insufficient illumination, limited color information, and low contrast. Summary of the Invention

[0006] In order to solve the above problems, the present invention proposes a method and system for night-time multi-target tracking based on domain adaptive learning. By performing image conversion on daytime data to simulate the night environment, guiding the student network based on the teacher network, learning the night data features in an unsupervised domain adaptive manner, and realizing night-time multi-target tracking, the method solves the problem of inaccurate night-time data tracking in the multi-target tracking problem and improves the multi-target tracking performance in the night environment.

[0007] According to some embodiments, a first solution of the present invention provides a method for nighttime multi-target tracking based on domain adaptive learning, which adopts the following technical solutions:

[0008] A nighttime multi-target tracking method based on domain adaptive learning, comprising:

[0009] Acquire daytime image data and nighttime image data;

[0010] Performing image conversion processing on the acquired daytime image data to obtain a nighttime conversion image;

[0011] Based on night image data, night conversion images and domain adaptive learning models, night multi-target tracking tasks are carried out to complete multi-target tracking in night environments;

[0012] Among them, the domain adaptive learning model adopts a teacher-student network, obtains tracking results based on the teacher network and night image data, and uses the obtained tracking results as pseudo labels; based on the pseudo labels, the student network is guided to perform unsupervised learning on night image data, and the student network performs supervised learning on night conversion images. The student network training is guided by the multi-domain learning loss value of the student network to obtain a domain adaptive learning model.

[0013] As a further technical limitation, night image data is input into the teacher network for image target tracking, the obtained preliminary image target tracking results are filtered by a confidence threshold, the preliminary image target tracking results exceeding the confidence threshold are used as tracking results, and the tracking results obtained by the teacher network are used to guide the pseudo-label training of the student network; the obtained tracking results are related to the weight value of the teacher network, and the teacher network weight can be obtained by an exponential moving average.

[0014] As a further technical limitation, unsupervised learning of the night image data by the student network is performed based on the obtained pseudo labels to obtain an unsupervised learning loss value.

[0015] Furthermore, the night-time conversion image is input into the student network for supervised learning of the image to obtain the supervised learning loss value; the obtained unsupervised learning loss value and supervised learning loss value are used to obtain the multi-domain learning loss value, and the student network is trained for multiple rounds according to the multi-domain learning loss value. When the multi-domain learning loss value is minimized, a domain adaptive learning model is obtained; according to the student network weight value obtained in each round of training, the exponential moving average is used to update the weight value of the teacher network.

[0016] As a further technical limitation, in the process of nighttime multi-target tracking, the acquired nighttime image data is input into the obtained domain adaptive model, and the output result of the domain adaptive model is the target tracking trajectory of the nighttime scene, thereby obtaining the nighttime multi-target tracking result.

[0017] As a further technical limitation, the image conversion processing includes at least image brightness adjustment, image contrast adjustment, image Gaussian noise removal and image Gaussian blur enhancement.

[0018] According to some embodiments, a second solution of the present invention provides a nighttime multi-target tracking system based on domain adaptive learning, which adopts the following technical solutions:

[0019] A nighttime multi-target tracking system based on domain adaptive learning, comprising:

[0020] an acquisition module configured to acquire daytime image data and nighttime image data;

[0021] a conversion module configured to perform image conversion processing on the acquired daytime image data to obtain a nighttime conversion image;

[0022] A tracking module is configured to perform nighttime multi-target tracking tasks based on nighttime image data, nighttime converted images, and a domain adaptive learning model to complete multi-target tracking in a nighttime environment;

[0023] Among them, in the tracking module, the domain adaptive learning model adopts a teacher-student network to obtain tracking results based on the teacher network and night image data, and uses the obtained tracking results as pseudo labels; based on the pseudo labels, the student network is guided to perform unsupervised learning on the night image data, and the student network performs supervised learning on the night conversion images. The student network training is guided by the multi-domain learning loss value of the student network to obtain a domain adaptive learning model.

[0024] According to some embodiments, a third solution of the present invention provides a computer-readable storage medium, which adopts the following technical solution:

[0025] A computer-readable storage medium stores a program, which, when executed by a processor, implements the steps of the nighttime multi-target tracking method based on domain adaptive learning as described in the first solution of the present invention.

[0026] According to some embodiments, a fourth solution of the present invention provides an electronic device, which adopts the following technical solution:

[0027] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the method for nighttime multi-target tracking based on domain adaptive learning as described in the first solution of the present invention.

[0028] According to some embodiments, a fifth solution of the present invention provides a computer program product, which adopts the following technical solution:

[0029] A computer program product includes software code, wherein the program in the software code executes the steps in the method for nighttime multi-target tracking based on domain adaptive learning as described in the first solution of the present invention.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] This embodiment simulates the nighttime environment by using daytime image data and performs supervised training of the domain adaptive learning model. This can effectively learn the data distribution of the nighttime environment and make full use of labeled daytime datasets and unlabeled nighttime datasets to train the domain adaptive learning model. This improves the performance of nighttime multi-target tracking while reducing the cost of nighttime data annotation. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The drawings constituting a part of the specification of this embodiment are used to provide a further understanding of this embodiment. The schematic embodiments and descriptions of this embodiment are used to explain this embodiment and do not constitute an improper limitation on this embodiment.

[0033] Figure 1 This is a basic flow chart of the nighttime multi-target tracking method based on domain adaptive learning in Example 1 of the present invention;

[0034] Figure 2 This is a training diagram of the domain adaptive learning model in the first embodiment of the present invention;

[0035] Figure 3 This is an architectural diagram of the nighttime multi-target tracking method based on domain adaptive learning in Example 1 of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0038] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0039] In the present invention, terms such as "upper", "lower", "left", "right", "front", "back", "vertical", "horizontal", "side", "bottom", etc. indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are relational words determined only for the convenience of describing the structural relationships of the various parts or elements of the present invention, and do not specifically refer to any part or element in the present invention, and should not be understood as limiting the present invention.

[0040] In the present invention, terms such as "fixed connection," "connected," and "connection" should be interpreted broadly to mean a fixed connection, an integral connection, or a detachable connection; a direct connection or an indirect connection through an intermediary. Relevant researchers or technicians in this field may determine the specific meanings of these terms in the present invention based on specific circumstances, and they should not be construed as limitations of the present invention.

[0041] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0042] Example 1

[0043] Embodiment 1 of the present invention introduces a method for nighttime multi-target tracking based on domain adaptive learning.

[0044] like Figure 1The figure shows a method for nighttime multi-target tracking based on domain adaptive learning. The method trains a domain adaptive learning model based on a labeled daytime dataset (in this embodiment, the domain adaptive learning model uses a multi-target tracking network), and then applies the trained multi-target tracking network to the nighttime dataset to obtain target tracking trajectories in nighttime scenes.

[0045] In order to realize multi-target tracking at night, this embodiment realizes multi-target tracking at night through a trained multi-target tracking network; the training process of the multi-target tracking network is as follows: Figure 2 As shown, the overall process steps are:

[0046] (1) Given a labeled daytime dataset and an unlabeled nighttime dataset, the multi-target tracking network is first trained using the daytime dataset to obtain the initialized multi-target tracking network weight parameters, and the initialized multi-target tracking network is used to construct the teacher-student model;

[0047] (2) Perform image conversion processing on the daytime dataset by random brightness, contrast, Gaussian noise, and Gaussian blur enhancement to obtain a labeled dataset with similar characteristics to the nighttime images;

[0048] (3) Inputting the image conversion dataset and the night dataset into the student network to obtain the tracking results; inputting the night dataset into the teacher network to obtain the tracking results, and filtering the tracking results through the confidence threshold to obtain high-confidence tracking results as pseudo labels to guide the training of the student model;

[0049] (4) The tracking result of the converted dataset output by the student network directly uses the labels of the daytime dataset to calculate the loss value of supervised learning. The supervised loss is calculated based on the results output by the student network and the labels of the daytime dataset:

[0050]

[0051] in, represents the classification loss, that is, the loss of the i-th anchor box being classified as the j-th true box category, represents the regression loss, that is, the loss of the i-th anchor box regressing to the j-th real box in terms of attributes such as position and size. λ represents a hyperparameter used to adjust the weight between classification loss and regression loss.

[0052]

[0053] L reg = -log(IoU(B gt ,B pred ))

[0054] Among them, n represents the total number of samples, i represents the index of the sample, representing the i-th sample, t i Represents the true label of the i-th sample, usually taking the value of 0 or 1, used to indicate the true category of the sample, p i Indicates the probability that the model predicts that the i-th sample is a positive class (such as category 1), B gt represents the ground-truth bounding box, B pred Represents the predicted bounding box, IoU represents the intersection ratio, and calculates the degree of overlap between the real box and the predicted box.

[0055] The output nighttime data set tracking results use the pseudo labels obtained by the teacher model to calculate the unsupervised learning loss value. The teacher network is used to guide the student network. The unsupervised loss is calculated as follows:

[0056]

[0057] Among them, KL represents the KL divergence function, represents the target category predicted by the teacher network, I t Represents the picture of the night dataset, C student represents the target category predicted by the student network.

[0058] The loss values ​​of supervised learning and unsupervised learning are used to guide the student network training, and the weight parameters of the teacher network are updated by the exponential moving average method. The update formula of the exponential moving average method is as follows:

[0059]

[0060] in, represents the teacher network parameters after the tth update, α is the smoothing factor, which usually takes a smaller value, and represents the weight of the new parameter update, θ t represents the student network parameters after the tth update, Represents the teacher network parameters after the previous update.

[0061] (5) Repeat the above steps (3) and (4) until the specified number of training rounds is reached.

[0062] This embodiment adopts Figure 2 The architecture diagram shown includes three modules: daytime image conversion, domain adaptive learning, and loss calculation. The input data of this embodiment includes a nighttime dataset and a daytime dataset for image conversion processing. After the input data, domain adaptive learning and loss calculation are performed to obtain the weight parameters of the multi-target tracking network in a nighttime environment. Then, the nighttime multi-target tracking task is performed to obtain the tracking trajectory of the target in the nighttime environment.

[0063] As one or more implementation methods, the obtained daytime dataset is subjected to image conversion processing, and the image conversion processing includes at least image brightness adjustment, image contrast adjustment, image Gaussian noise removal, and image Gaussian blur enhancement; the obtained image conversion data is used to simulate the data characteristics of nighttime images, making the multi-target tracking network training in domain adaptive learning more robust.

[0064] In one or more implementations, the domain adaptive learning module utilizes a teacher-student network to train a multi-target tracking network in a nighttime environment. The student network is primarily used for multi-domain training on an image-converted daytime dataset and an unlabeled nighttime dataset. The teacher network is used to generate pseudo-labels on the unlabeled nighttime dataset and provide them to the student model. The weights of the pseudo-labels are obtained by the student model using an exponential moving average (EMA). To generate more reliable pseudo-labels, the tracking results output by the teacher network are filtered using a confidence threshold to obtain high-confidence tracking results as the final pseudo-labels to guide the training of the student network.

[0065] In one or more embodiments, the calculated loss includes supervised learning and unsupervised learning. Supervised learning is performed on the data obtained after image conversion processing, and the loss between the tracking results and the labels is calculated. Furthermore, the unsupervised loss of the student network under the guidance of the teacher network's pseudo-labels is calculated. The parameters of the multi-target tracking network are continuously adjusted based on the calculated loss values ​​to train the network so that the tracking results are as close to the real results as possible.

[0066] As one or more implementation methods, during the process of nighttime multi-target tracking, the acquired nighttime image data is input into the obtained domain adaptive model, and the output result of the domain adaptive model is the target tracking trajectory of the nighttime scene, thereby obtaining a nighttime multi-target tracking result.

[0067] Due to the significant differences between daytime and nighttime data, a multi-target tracking network trained on a labeled daytime dataset will experience significant performance degradation when directly applied to a nighttime environment. Therefore, this embodiment adds two additional steps: image conversion processing and domain adaptation learning. Image conversion processing is first performed on the labeled daytime dataset to simulate a nighttime environment. Due to the large gap between nighttime data and daytime data in terms of illumination, color information, contrast, image noise, etc., and the lack of label information in the nighttime dataset, training the multi-target tracking network using only the daytime dataset cannot achieve good results in the nighttime environment; using the daytime dataset after image conversion processing can better simulate the characteristics of nighttime images and extract the robust features of the nighttime environment and targets; after the image conversion processing is completed, the multi-target tracking network is trained using the teacher-student model in domain adaptive learning. The student network in the teacher-student model uses the labeled daytime dataset and the unlabeled nighttime dataset after image conversion processing as input, while the teacher network only uses the nighttime dataset as input to generate pseudo labels to guide the student network; in domain adaptive learning, the teacher-student model is used to perform supervised learning on the converted daytime data and unsupervised learning on the unlabeled nighttime data, which can fully utilize all daytime and nighttime data to train the multi-target tracking network, effectively solving the distribution difference between daytime and nighttime data, thereby improving the multi-target tracking performance in the nighttime environment.

[0068] This embodiment is based on the simulation of nighttime data converted from daytime images. The daytime dataset is processed through random brightness, contrast, Gaussian noise, and Gaussian blur enhancement to obtain a labeled dataset with similar characteristics to nighttime images. The teacher-student model is used for unsupervised domain adaptive learning to improve the tracking performance of the multi-target tracking network in unlabeled nighttime datasets.

[0069] This embodiment performs image conversion processing on the daytime dataset so that the daytime dataset can more accurately simulate the nighttime environment. By using this dataset simulating the nighttime environment instead of the original daytime dataset for supervised training of the multi-target tracking network, the data distribution of the nighttime environment can be effectively learned. At the same time, in order to make full use of a large amount of unlabeled nighttime data, this embodiment uses a teacher-student model for unsupervised domain adaptive learning, wherein the teacher network takes the unlabeled nighttime dataset as input, filters the output tracking results through a confidence threshold, and obtains the filtered high-confidence results as pseudo-labels to guide the student network for unsupervised learning. The entire domain adaptive learning process makes full use of the labeled daytime dataset and the unlabeled nighttime dataset to train the multi-target tracking network, thereby improving the performance of multi-target tracking at night while reducing the cost of nighttime data labeling.

[0070] Example 2

[0071] The second embodiment of the present invention introduces a nighttime multi-target tracking system based on domain adaptive learning.

[0072] A nighttime multi-target tracking system based on domain adaptive learning, comprising:

[0073] an acquisition module configured to acquire daytime image data and nighttime image data;

[0074] a conversion module configured to perform image conversion processing on the acquired daytime image data to obtain a nighttime conversion image;

[0075] A tracking module is configured to perform nighttime multi-target tracking tasks based on nighttime image data, nighttime converted images, and a domain adaptive learning model to complete multi-target tracking in a nighttime environment;

[0076] Among them, in the tracking module, the domain adaptive learning model adopts a teacher-student network to obtain tracking results based on the teacher network and night image data, and uses the obtained tracking results as pseudo labels; based on the pseudo labels, the student network is guided to perform unsupervised learning on the night image data, and the student network performs supervised learning on the night conversion images. The student network training is guided by the multi-domain learning loss value of the student network to obtain a domain adaptive learning model.

[0077] The detailed steps are the same as those of the nighttime multi-target tracking method based on domain adaptive learning provided in Example 1 and will not be repeated here.

[0078] Example 3

[0079] A third embodiment of the present invention provides a computer-readable storage medium.

[0080] A computer-readable storage medium stores a program, which, when executed by a processor, implements the steps of the method for nighttime multi-target tracking based on domain adaptive learning as described in the first embodiment of the present invention.

[0081] The detailed steps are the same as those of the nighttime multi-target tracking method based on domain adaptive learning provided in Example 1 and will not be repeated here.

[0082] Example 4

[0083] A fourth embodiment of the present invention provides an electronic device.

[0084] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps of the method for nighttime multi-target tracking based on domain adaptive learning as described in Example 1 of the present invention.

[0085] The detailed steps are the same as those of the nighttime multi-target tracking method based on domain adaptive learning provided in Example 1 and will not be repeated here.

[0086] Example 5

[0087] A fifth embodiment of the present invention provides a computer program product.

[0088] A computer program product includes software code, wherein the program in the software code executes the steps of the method for nighttime multi-target tracking based on domain adaptive learning as described in the first embodiment of the present invention.

[0089] The detailed steps are the same as those of the nighttime multi-target tracking method based on domain adaptive learning provided in Example 1 and will not be repeated here.

[0090] The above description is merely a preferred embodiment of this embodiment and is not intended to limit this embodiment. Those skilled in the art will readily appreciate that this embodiment may be modified and varied in various ways. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this embodiment shall be within the scope of protection of this embodiment.

Claims

1. A nighttime multi-target tracking method based on domain adaptive learning, characterized in that: include: Acquire daytime image data and nighttime image data; Performing image conversion processing on the acquired daytime image data to obtain a nighttime converted image, wherein the image conversion processing at least includes image brightness adjustment, image contrast adjustment, image Gaussian noise removal, and image Gaussian blur enhancement; Based on night image data, night conversion images and domain adaptive learning models, night multi-target tracking tasks are carried out to complete multi-target tracking in night environments; The domain adaptive learning model uses a teacher-student network. The student network in the teacher-student model uses a labeled daytime dataset and an unlabeled nighttime dataset after image conversion as input, while the teacher network only uses the nighttime dataset as input to generate pseudo-labels to guide the student network. Tracking results are obtained based on the teacher network and the nighttime image data, and the obtained tracking results are used as pseudo-labels. The pseudo-labels guide the student network to perform unsupervised learning on the nighttime image data, while the student network performs supervised learning on the converted nighttime images. The student network is trained based on its multi-domain learning loss value to obtain a domain adaptive learning model. The output nighttime data set tracking results use the pseudo labels obtained by the teacher model to calculate the unsupervised learning loss value. The teacher network is used to guide the student network. The unsupervised loss is calculated as follows: in, represents the classification loss, that is, the loss of the i-th anchor box being classified as the j-th true box category, Represents the regression loss, that is, the loss of the i-th anchor box regressing to the j-th real box in terms of attributes such as position and size, Represents a hyperparameter used to adjust the weight between classification loss and regression loss; represents the KL divergence function, represents the target category predicted by the teacher network, Represents a picture of the nighttime dataset, represents the target category predicted by the student network.

2. A method for nighttime multi-target tracking based on domain adaptive learning as claimed in claim 1, characterized in that: Nighttime image data is input into a teacher network for image target tracking, and the obtained preliminary image target tracking results are filtered by a confidence threshold. The preliminary image target tracking results exceeding the confidence threshold are used as tracking results, and the tracking results obtained by the teacher network are used to guide the pseudo-label training of the student network; the obtained tracking results are related to the weight value of the teacher network, and the teacher network weight can be obtained by an exponential moving average.

3. The method for nighttime multi-target tracking based on domain adaptive learning as claimed in claim 1, characterized in that: The student network is used to perform unsupervised learning on the night image data based on the obtained pseudo labels to obtain the unsupervised learning loss value.

4. A method for nighttime multi-target tracking based on domain adaptive learning as claimed in claim 3, characterized in that: The nighttime conversion image is input into the student network for supervised learning of the image to obtain the supervised learning loss value; the obtained unsupervised learning loss value and supervised learning loss value are used to obtain the multi-domain learning loss value, and the student network is trained multiple rounds according to the multi-domain learning loss value. When the multi-domain learning loss value is minimized, a domain adaptive learning model is obtained; according to the student network weight value obtained in each round of training, the exponential moving average is used to update the weight value of the teacher network.

5. The method for nighttime multi-target tracking based on domain adaptive learning as claimed in claim 1, characterized in that: In the process of nighttime multi-target tracking, the acquired nighttime image data is input into the obtained domain adaptive model, and the output result of the domain adaptive model is the target tracking trajectory of the nighttime scene, thereby obtaining the nighttime multi-target tracking result.

6. The method for nighttime multi-target tracking based on domain adaptive learning as claimed in claim 1, characterized in that: The image conversion process at least includes image brightness adjustment, image contrast adjustment, image Gaussian noise removal and image Gaussian blur enhancement.

7. A nighttime multi-target tracking system based on domain adaptive learning, characterized in that: include: an acquisition module configured to acquire daytime image data and nighttime image data; a conversion module configured to perform image conversion processing on the acquired daytime image data to obtain a nighttime converted image, wherein the image conversion processing at least includes image brightness adjustment, image contrast adjustment, image Gaussian noise removal, and image Gaussian blur enhancement; A tracking module is configured to perform nighttime multi-target tracking tasks based on nighttime image data, nighttime converted images, and a domain adaptive learning model to complete multi-target tracking in a nighttime environment; In the tracking module, a domain adaptive learning model uses a teacher-student network. The student network in the teacher-student model uses a labeled daytime dataset and an unlabeled nighttime dataset after image transformation as input, while the teacher network uses only the nighttime dataset as input to generate pseudo-labels to guide the student network. Tracking results are obtained based on the teacher network and the nighttime image data, and the obtained tracking results are used as pseudo-labels. The pseudo-labels guide the student network to perform unsupervised learning on the nighttime image data, while the student network performs supervised learning on the converted nighttime images. The student network is trained based on its multi-domain learning loss value, resulting in a domain adaptive learning model. The output nighttime data set tracking results use the pseudo labels obtained by the teacher model to calculate the unsupervised learning loss value. The teacher network is used to guide the student network. The unsupervised loss is calculated as follows: in, represents the classification loss, that is, the loss of the i-th anchor box being classified as the j-th true box category, Represents the regression loss, that is, the loss of the i-th anchor box regressing to the j-th real box in terms of attributes such as position and size, Represents a hyperparameter used to adjust the weight between classification loss and regression loss; represents the KL divergence function, represents the target category predicted by the teacher network, Represents a picture of the nighttime dataset, represents the target category predicted by the student network.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the nighttime multi-target tracking method based on domain adaptive learning are implemented as described in any one of claims 1 to 6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the program, the steps of the nighttime multi-target tracking method based on domain adaptive learning are implemented as described in any one of claims 1 to 6.

10. A computer program product comprising software code, characterized in that The program in the software code executes the steps of the nighttime multi-target tracking method based on domain adaptive learning according to any one of claims 1 to 6.

Images