Knowledge distillation method and system for target tracking

Abstract

The invention provides a knowledge distillation method and system for target tracking, and relates to the technical field of target tracking. Extracting space attention and channel attention from the feature fusion layer of the teacher model through mask-based feature distillation, generating an attention mask, and multiplying the attention mask by the feature map of the student model; based on a context adaptive learning mode, adaptively integrating space and feature knowledge of each layer of the teacher model through a multi-layer perceptron so that the student model can adaptively adjust learning key points according to different depths of the network; and carrying out distillation through correlation distance supervision, projecting regression frame prediction of the teacher model and the student model into a probability distribution space by adopting normal affine transformation, and supervising an ability gap between the teacher model and the student model by evaluating a correlation distance of probability distribution. The technical effects that the universality is improved in the target tracking process, and the performance is superior to that of an existing baseline are achieved.

Description

Technical Field

[0001] The present application relates to the field of tracking video technology, and in particular to a knowledge distillation method and system for target tracking. Background Art

[0002] Video is a typical form of multimedia data. Tracking important objects in video content is a critical task, and with the development of smart embedded devices, object tracking through edge devices is becoming increasingly important in various fields, including intelligent transportation systems and disaster relief operations. This mitigates the potential delays associated with transmitting large amounts of data, thereby facilitating more efficient capture of time-sensitive targets. In addition, it helps tracking models operate effectively in situations with limited bandwidth and unstable communications.

[0003] Currently, there are two main trends in the field of automatic real-time object tracking on edge devices. The first approach focuses on simplifying the model at the structural level, which includes compressing large models and designing lightweight alternatives. However, this compensation often leads to a significant drop in performance as the compression ratio increases. On the other hand, developing network architectures that are both lightweight and high-performance remains a considerable challenge. In contrast, another approach emphasizes the use of knowledge distillation (KD) techniques to integrate the knowledge gained from large models into smaller models. This approach has an advantage over previous methods in that it can provide more accurate adaptive supervision information. However, some challenges still exist when adapting KD to object tracking on edge devices.

[0004] Currently, there are two dominant tracking model architectures: Siamese network-based and Vit-based, and their tracking pipelines are significantly different. In addition, for their respective backbone networks: Convolutional Neural Networks (CNN) and Visual Transformers (ViT), existing studies have shown that the KD algorithm that favors CNN may have an adverse effect on ViT. Therefore, it is urgent to study knowledge distillation methods that can be effectively applied to both tracking architectures.

[0005] Second, the object tracking task is divided into foreground and background regions only. This problem is similar to the binary classification problem, which limits the enhancement through logit distillation. In terms of feature extraction, the background may contain distracting objects that are similar to the target or belong to the same category. However, most existing methods ignore the inherent conflict between classification and regression.

[0006] Third, a significant challenge associated with current KD techniques applied to tracking models is that the benefits gained from KD are significantly reduced when the difference between the teacher and student models is too significant. A common strategy to mitigate this problem involves adding an intermediate teacher model to facilitate the transition, which significantly increases the training overhead.

[0007] Therefore, how to improve the versatility in the target tracking process and outperform the existing baseline performance has become a technical problem that needs to be solved urgently. Summary of the invention

[0008] In order to achieve accurate prediction of carbon prices, the present application provides a knowledge distillation method and system for target tracking.

[0009] In the first aspect, the present application provides a knowledge distillation method for target tracking using the following technical solution: A knowledge distillation method for target tracking, comprising: Mask-based feature distillation extracts spatial attention and channel attention from the feature fusion layer of the teacher model, generates an attention mask, and multiplies the attention mask with the feature map of the student model; Based on context-adaptive learning, the spatial and feature knowledge of each layer of the teacher model is adaptively integrated through a multi-layer perceptron so that the student model can adaptively adjust the learning focus according to the different depths of the network; Distillation is performed through correlation distance supervision. A positive affine transformation is used to project the regression box predictions of the teacher model and the student model into the probability distribution space, and the ability gap between the teacher model and the student model is supervised by evaluating the correlation distance of the probability distribution.

[0010] Optionally, the mask-based feature distillation, extracting spatial attention and channel attention from the feature fusion layer of the teacher model, generating an attention mask, and multiplying the attention mask with the feature map of the student model, comprises: Obtain the teacher model and the student model, align the feature maps of the teacher model and the student model, and ensure that the size of the feature map of the student model is consistent with that of the teacher model; Extracting spatial attention and channel attention from the feature fusion layer of the teacher model to generate an attention mask; The attention mask is multiplied with the feature maps of the teacher model and the student model.

[0011] Optionally, the context-based adaptive learning method adaptively integrates the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron so that the student model can adaptively adjust the learning focus according to the different depths of the network, including: Based on context-adaptive learning, the feature maps of the teacher model and the student model are pooled at multiple scales to be divided into shallow, medium and deep features. Adaptively integrate the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron; The learning focus of the student model is adjusted so that the student model can be adaptively adjusted according to the spatial and feature knowledge of each layer in the teacher model.

[0012] Optionally, the step of performing distillation by correlation distance supervision, projecting the regression box predictions of the teacher model and the student model into the probability distribution space by using a positive affine transformation, and supervising the ability gap between the teacher model and the student model by evaluating the correlation distance of the probability distribution, comprises: Pearson distance is used as a metric to evaluate the correlation distance between the teacher model and the student model in the probability distribution space; Combining the correlation distance and using a positive affine transformation to project the regression box predictions of the teacher model and the student model into a probability distribution space; The ability gap between the teacher model and the student model is supervised by the inter-class position correlation and the intra-class position correlation in the probability distribution space.

[0013] Optionally, the method is applicable to a Siamese network-based and a visual Transformer-based target tracking model.

[0014] Optionally, the method enhances the student model's ability to focus on the target area and reduces the impact of low resolution and intra-class interference by combining spatial attention and channel attention.

[0015] Optionally, the method reduces overfitting of the student model to the teacher model output through relevant distance supervision, thereby improving the tracking performance of the student model in complex scenarios.

[0016] In a second aspect, the present application provides a knowledge distillation system for target tracking, the knowledge distillation system for target tracking comprising: An attention mask generation module for mask-based feature distillation, extracting spatial attention and channel attention from the feature fusion layer of the teacher model, generating an attention mask, and multiplying the attention mask with the feature map of the student model; The learning adjustment module is used to adaptively integrate the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron based on context-based adaptive learning so that the student model can adaptively adjust the learning focus according to the different depths of the network; The capability gap assessment module, used for distillation through correlation distance supervision, adopts positive affine transformation to project the regression box predictions of the teacher model and the student model into the probability distribution space, and supervises the capability gap between the teacher model and the student model by evaluating the correlation distance of the probability distribution.

[0017] In a third aspect, the present application provides a computer device, comprising: a memory and a processor, wherein the processor executes the method described above when running computer instructions stored in the memory.

[0018] In a fourth aspect, the present application provides a computer-readable storage medium, comprising instructions, which, when executed on a computer, enable the computer to execute the method described above.

[0019] In summary, the present application includes the following beneficial technical effects: This application extracts spatial attention and channel attention from the feature fusion layer of the teacher model through mask-based feature distillation, generates an attention mask, and multiplies the attention mask with the feature map of the student model; based on the context-adaptive learning method, the spatial and feature knowledge of each layer of the teacher model is adaptively integrated through a multi-layer perceptron so that the student model can adaptively adjust the learning focus according to the different depths of the network; distillation is performed through relevant distance supervision, and the regression box predictions of the teacher model and the student model are projected into the probability distribution space using a positive affine transformation, and the ability gap between the teacher model and the student model is supervised by evaluating the relevant distance of the probability distribution. The technical effect of improving versatility in the process of target tracking and outperforming the performance of the existing baseline is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiment of the present application; Figure 2 It is a flowchart of an embodiment of a knowledge distillation method for target tracking in the present application; Figure 3 This is a conceptual diagram of the process in the knowledge distillation method used for target tracking in this application; Figure 4 It is a feature conversion graph attached to the VIT-based model in the knowledge distillation method for target tracking in this application; Figure 5 It is a structural block diagram of an embodiment of a knowledge distillation system for target tracking in the present application. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below through the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0022] Reference Figure 1 , Figure 1 A schematic diagram of the computer device structure of the hardware operating environment involved in the embodiment of the present application.

[0023] like Figure 1 As shown, the computer device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM), or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk storage. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0024] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the computer device, and may include more or less components than shown in the figure, or combine some components, or arrange the components differently.

[0025] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and a knowledge distillation program for target tracking.

[0026] exist Figure 1 In the computer device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the present application can be set in the computer device, and the computer device calls the knowledge distillation program for target tracking stored in the memory 1005 through the processor 1001, and executes the knowledge distillation method for target tracking provided in the embodiment of the present application.

[0027] The present application embodiment provides a knowledge distillation method for target tracking. Figure 2 , Figure 2 A flowchart of an embodiment of a knowledge distillation method for target tracking in the present application.

[0028] In this embodiment, the knowledge distillation method for target tracking includes the following steps: Step S10: Mask-based feature distillation, extracting spatial attention and channel attention from the feature fusion layer of the teacher model, generating an attention mask, and multiplying the attention mask with the feature map of the student model.

[0029] It should be noted that the steps of extracting spatial attention and channel attention from the feature fusion layer of the teacher model, generating an attention mask, and multiplying the attention mask with the feature map of the student model based on mask feature distillation include: obtaining the teacher model and the student model, aligning the feature maps of the teacher model and the student model, and ensuring that the feature map of the student model is of the same size as the feature map of the teacher model; extracting spatial attention and channel attention from the feature fusion layer of the teacher model to generate an attention mask; and multiplying the attention mask with the feature maps of the teacher model and the student model.

[0030] In the specific implementation, this example provides a comprehensive description of the OTKD pipeline related to the target tracking task. It consists of three stages, such as Figure 3 In the context of the viti-based model, the process consists of executing Figure 4 The operations shown in . The features are extracted and classified into four different categories. Subsequently, these features are reconstructed into a two-dimensional format and the OTKD method is applied. Figure 3 The curve in the figure is the probability distribution of the four values ​​of the regression box. The coordinates can be expressed as x and y (each frame of each video sequence has these four values, which will form a probability distribution after being projected into two-dimensional space. The function curve is just a fitting representation of the probability distribution).

[0031] In specific implementation, compared with tasks such as target classification or detection, tracking tasks only distinguish between background and foreground areas. However, there are interference elements in the background that are similar to the target, which requires the tracker to have both feature extraction and region prediction capabilities (corresponding to classification and regression, respectively).

[0032] Knowledge is extracted from the feature fusion layer of the teacher model through multi-layer feature distillation. First, the feature maps of each layer in the teacher model are fixed. Then, the feature maps of the corresponding layers in the student model are aligned to ensure that their size is consistent with the teacher model. Subsequently, the dimensions of the deep feature maps in the student model are modified to match the structure of the shallow feature maps.

[0033] On this basis, mask-based features (MBF) are developed to extract spatial attention and channel attention from the feature maps generated by each layer of the teacher model. Assume that the feature fusion layers of the teacher model are m1T, m2T, …, mNT. N is the number of feature fusion layers. The student model is m1S, m2S, …, mNs. These two forms of attention are used as attention masks and multiplied with each feature map respectively. Subsequently, the features after the above operations are combined to form the final output. The two attention masks can be described as: in, represents the channel attention mask of the i-th layer of the student model, represents the spatial attention mask of the i-th layer of the student model, is the feature map of the i-th layer of the student model, , , Respectively represent the number, height and width of the channels of the i-th layer of the student model. Among them, Characteristic graph representing students Feature map of the adjusted and teacher model The feature map after size alignment. is the loss after channel attention mask, is the loss after the spatial attention mask. D represents the distance function used to quantify the difference between the student and teacher models, and ⊙ represents the multiplication broadcast operation.

[0034] Step S20: Based on the context-adaptive learning approach, the spatial and feature knowledge of each layer of the teacher model is adaptively integrated through a multi-layer perceptron so that the student model can adaptively adjust the learning focus according to the different depths of the network.

[0035] It can be understood that the steps of adaptively integrating the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron based on context adaptive learning so that the student model can adaptively adjust the learning focus according to the different depths of the network include: multi-scale pooling of the feature maps of the teacher model and the student model based on context adaptive learning to be divided into shallow, middle and deep features; adaptively integrating the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron; adjusting the learning focus of the student model so that the student model can adaptively adjust according to the spatial and feature knowledge of each layer in the teacher model.

[0036] It should be noted that through the above mechanism, the student model can benefit from the guidance provided by each level of the teacher model. However, in the process of matching transformation and multi-layer fusion, the hierarchical information of the feature map is destroyed. On the other hand, although this study uses the MBF method to divide knowledge into spatial information knowledge and target appearance knowledge, there are differences in the importance of feature information and spatial information in neural networks at different levels. This embodiment introduces a contextual adaptive learning mechanism (CAL). When solving the first problem, the feature map is pooled at three scales, divided into shallow, medium and deep, denoted as P(). Therefore, the knowledge is divided into different levels of context information.

[0037] Before multi-scale pooling, a three-layer MLP is used to adaptively integrate the spatial and feature knowledge of each layer of the teacher model. This gives the student model enhanced adaptive adjustment capabilities. Subsequently, the L2 loss between the teacher model and the student model is calculated. The above comprehensive loss can be converted into the following form: in, is the feature of the jth layer of the student model after the channel attention mask and the spatial attention mask, is the feature of the jth layer of the teacher model after the channel attention mask and the spatial attention mask. P() is the pooling function. MLP() is the multi-layer perceptron operation. , , It is the feature of three scales after multi-scale pooling of the masked feature of the jth layer of the teacher model , , It is the features of three scales after multi-scale pooling of the masked features of the j-th layer of the student model. is the loss function.

[0038] This embodiment incorporates the parameters of the MLP into the iterative training of the tracking model, aiming to enable the student model to autonomously determine whether the focus of learning at different scales and levels of the network is spatial knowledge or feature knowledge.

[0039] Step S30: Distillation is performed through correlation distance supervision, using a positive affine transformation to project the regression box predictions of the teacher model and the student model into the probability distribution space, and the ability gap between the teacher model and the student model is supervised by evaluating the correlation distance of the probability distribution.

[0040] In a specific implementation, distillation is performed through correlation distance supervision, a positive affine transformation is used to project the regression box predictions of the teacher model and the student model into a probability distribution space, and the ability gap between the teacher model and the student model is supervised by evaluating the correlation distance of the probability distribution, including: using the Pearson distance as a metric to evaluate the correlation distance between the teacher model and the student model in the probability distribution space; combining the correlation distance and using a positive affine transformation to project the regression box predictions of the teacher model and the student model into the probability distribution space; supervising the ability gap between the teacher model and the student model in the probability distribution space through inter-class position correlation and intra-class position correlation.

[0041] It should be noted that the method is applicable to target tracking models based on twin networks and visual Transformer.

[0042] It can be understood that the method enhances the student model's ability to focus on the target area and reduces the impact of low resolution and intra-class interference by combining spatial attention and channel attention. In the specific implementation, the method reduces the overfitting of the student model to the teacher model output through relevant distance supervision, and improves the tracking performance of the student model in complex scenarios.

[0043] It should be noted that the regression box depicts the teacher's prediction of the target position in each frame. Classical KD requires an exact correspondence between the teacher's outputs for each frame, and the best performance can only be achieved when all values ​​appear completely consistent. This brings two related problems. First, the student model exhibits a high sensitivity to overfitting, resulting in suboptimal generalization in specific cases. This phenomenon will be further clarified by subsequent analysis of attention visualization. On the other hand, the regression bounding boxes generated by the teacher model are distributed in a discrete manner in space, which poses a challenge to the student model to determine the optimal fitting plane. Therefore, this embodiment uses a positive affine transformation to project the video sequence into a probability distribution space. It supervises the ability gap between teachers and students by evaluating the relevant distance of the probability distribution in a specified space, and calls this concept inter-class position correlation.

[0044] In the specific implementation, similar or distracting features of the same category may also appear in each video sequence. The teacher model shows different confidence levels in different regions. The relationship between these confidence levels is represented as intra-class position correlation, which can then be transferred to the student model: represents the Pearson correlation coefficient that quantifies the teacher-student relationship in the i-th video sequence, is the number of picture frames in the i-th video sequence.

[0045] Therefore, the total training loss can be expressed as: . represents the label loss, and , and is the coefficient of the balance loss component.

[0046] It should be noted that this embodiment proposes a new knowledge distillation method for target tracking, called OTKD, which is applicable to two mainstream tracking architectures. The method of this embodiment combines attention masks to promote the distillation of feature knowledge and introduces MLP to achieve the learning ability of adaptive distillation. In addition, it also guides students to focus on understanding the probability distribution of the regression boxes generated by the teacher model, rather than striving to accurately correspond to the teacher's output.

[0047] In this embodiment, one of the implementation environments is: the input is a video stream captured by an edge device such as a drone and the initial position of the target to be tracked (manually framed or initially framed by target detection), thereby inputting the target image and the search area image into the tracking model. The tracking model extracts the features of the target image and the search area image respectively and compares them, and selects the area most relevant to the target (the appearance features are most similar, and the spatial position features should maintain temporal sequence) as the next frame position of the target tracking, and outputs the regression box coordinates of the position (the algorithm automatically frames).

[0048] This embodiment extracts spatial attention and channel attention from the feature fusion layer of the teacher model through mask-based feature distillation, generates an attention mask, and multiplies the attention mask with the feature map of the student model; based on the context-adaptive learning method, the spatial and feature knowledge of each layer of the teacher model is adaptively integrated through a multi-layer perceptron so that the student model can adaptively adjust the learning focus according to the different depths of the network; distillation is performed through relevant distance supervision, and the regression box predictions of the teacher model and the student model are projected into the probability distribution space using a positive affine transformation, and the ability gap between the teacher model and the student model is supervised by evaluating the relevant distance of the probability distribution. The technical effect of improving versatility in the process of target tracking and outperforming the performance of the existing baseline is achieved.

[0049] In addition, an embodiment of the present application also proposes a computer-readable storage medium, on which a program for knowledge distillation for target tracking is stored. When the program for knowledge distillation for target tracking is executed by a processor, the steps of the method for knowledge distillation for target tracking as described above are implemented.

[0050] Reference Figure 5 , Figure 5 This is a structural block diagram of an embodiment of a knowledge distillation system for target tracking in the present application.

[0051] like Figure 5 As shown, the knowledge distillation system for target tracking proposed in the embodiment of the present application includes: An attention mask generation module 10, for extracting spatial attention and channel attention from the feature fusion layer of the teacher model based on mask feature distillation, generating an attention mask, and multiplying the attention mask with the feature map of the student model; A learning adjustment module 20, which is used to adaptively integrate the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron based on context-based adaptive learning so that the student model can adaptively adjust the learning focus according to different depths of the network; The capability gap assessment module 30 is used for distillation through correlation distance supervision, which projects the regression box predictions of the teacher model and the student model into the probability distribution space using a positive affine transformation, and supervises the capability gap between the teacher model and the student model by evaluating the correlation distance of the probability distribution.

[0052] It should be understood that the above is only an example and does not constitute any limitation on the technical solution of the present application. In specific applications, technicians in this field can make settings as needed, and the present application does not impose any limitation on this.

[0053] This embodiment extracts spatial attention and channel attention from the feature fusion layer of the teacher model through mask-based feature distillation, generates an attention mask, and multiplies the attention mask with the feature map of the student model; based on the context-adaptive learning method, the spatial and feature knowledge of each layer of the teacher model is adaptively integrated through a multi-layer perceptron so that the student model can adaptively adjust the learning focus according to the different depths of the network; distillation is performed through relevant distance supervision, and the regression box predictions of the teacher model and the student model are projected into the probability distribution space using a positive affine transformation, and the ability gap between the teacher model and the student model is supervised by evaluating the relevant distance of the probability distribution. The technical effect of improving versatility in the process of target tracking and outperforming the performance of the existing baseline is achieved.

[0054] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present application. In practical applications, technicians in this field can select part or all of it according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0055] In addition, for technical details not described in detail in this embodiment, please refer to the knowledge distillation method for target tracking provided in any embodiment of the present application, which will not be repeated here.

[0056] In addition, it should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0057] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0058] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory (ROM) / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods of each embodiment of the present application.

[0059] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A knowledge distillation method for target tracking, characterized in that: include: Mask-based feature distillation extracts spatial attention and channel attention from the feature fusion layer of the teacher model, generates an attention mask, and multiplies the attention mask with the feature map of the student model; Based on context-adaptive learning, the spatial and feature knowledge of each layer of the teacher model is adaptively integrated through a multi-layer perceptron so that the student model can adaptively adjust the learning focus according to the different depths of the network; Distillation is performed through correlation distance supervision. A positive affine transformation is used to project the regression box predictions of the teacher model and the student model into the probability distribution space, and the ability gap between the teacher model and the student model is supervised by evaluating the correlation distance of the probability distribution.

2. The knowledge distillation method for target tracking according to claim 1, characterized in that: The mask-based feature distillation extracts spatial attention and channel attention from the feature fusion layer of the teacher model, generates an attention mask, and multiplies the attention mask with the feature map of the student model, including: Obtain the teacher model and the student model, align the feature maps of the teacher model and the student model, and ensure that the size of the feature map of the student model is consistent with that of the teacher model; Extracting spatial attention and channel attention from the feature fusion layer of the teacher model to generate an attention mask; The attention mask is multiplied with the feature maps of the teacher model and the student model.

3. The knowledge distillation method for target tracking according to claim 1, characterized in that: The context-based adaptive learning method adaptively integrates the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron so that the student model can adaptively adjust the learning focus according to the different depths of the network, including: Based on context-adaptive learning, the feature maps of the teacher model and the student model are pooled at multiple scales to be divided into shallow, medium and deep features. Adaptively integrate the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron; The learning focus of the student model is adjusted so that the student model can be adaptively adjusted according to the spatial and feature knowledge of each layer in the teacher model.

4. The knowledge distillation method for target tracking according to claim 1, characterized in that: The step of performing distillation by correlation distance supervision, projecting the regression box predictions of the teacher model and the student model into the probability distribution space by using a positive affine transformation, and supervising the ability gap between the teacher model and the student model by evaluating the correlation distance of the probability distribution, includes: Pearson distance is used as a metric to evaluate the correlation distance between the teacher model and the student model in the probability distribution space; Combining the correlation distance and using a positive affine transformation to project the regression box predictions of the teacher model and the student model into a probability distribution space; The ability gap between the teacher model and the student model is supervised by the inter-class position correlation and the intra-class position correlation in the probability distribution space.

5. The knowledge distillation method for target tracking according to claim 1, characterized in that: The method is applicable to both Siamese network-based and visual Transformer-based object tracking models.

6. The knowledge distillation method for target tracking according to claim 1, characterized in that: The method combines spatial attention and channel attention to enhance the student model's ability to focus on the target area and reduce the impact of low resolution and intra-class interference.

7. The knowledge distillation method for target tracking according to claim 1, characterized in that: The method reduces the overfitting of the student model to the teacher model output through relevant distance supervision, thereby improving the tracking performance of the student model in complex scenarios.

8. A knowledge distillation system for target tracking, characterized in that: The knowledge distillation system for target tracking includes: An attention mask generation module for mask-based feature distillation, extracting spatial attention and channel attention from the feature fusion layer of the teacher model, generating an attention mask, and multiplying the attention mask with the feature map of the student model; The learning adjustment module is used to adaptively integrate the spatial and feature knowledge of each layer of the teacher model through a multi-layer perceptron based on context-based adaptive learning so that the student model can adaptively adjust the learning focus according to the different depths of the network; The capability gap assessment module, used for distillation through correlation distance supervision, adopts positive affine transformation to project the regression box predictions of the teacher model and the student model into the probability distribution space, and supervises the capability gap between the teacher model and the student model by evaluating the correlation distance of the probability distribution.

9. A computer device, characterized in that: The device comprises: a memory and a processor, and when the processor runs the computer instructions stored in the memory, the processor executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The method comprises instructions, which, when executed on a computer, cause the computer to execute the method according to any one of claims 1 to 7.

Images