Target tracking method based on deep learning twin network

Abstract

The invention relates to a method in the field of target tracking, in particular to a target tracking method based on a deep learning twin network, and aims to solve the problems that in the prior art, a tracking model can generate pollution superposition under the scenes of target shielding, severe deformation, view field exiting and the like; the difficulty of distinguishing similar targets is high; and a disordered background is easy to pollute the similar targets. The target tracking method based on the deep learning twin network comprises the following steps: capturing local and global context information by using a deep separable convolutional network, enhancing features by using a multi-scale large-kernel module, carrying out association operation by using the twin network, and obtaining a target bounding box by using a classification regression network, so that target tracking is accurate, efficient and flexible; compared with a traditional correlation filter algorithm, the method has better performance and algorithm robustness.

Description

Technical Field

[0001] The present invention relates to target tracking methods, and in particular to a target tracking method based on a deep learning twin network. Background Art

[0002] Video target tracking is one of the important basic research issues in the field of computer vision. Video target tracking means that after the target is specified in the first frame of the video sequence (the target bounding box is given, usually represented by a rectangular box), the target specified in the first frame is continuously tracked in subsequent frames, and the target is calibrated using the target bounding box to achieve the positioning and scale estimation of the target. Video target tracking plays a very important role in fields such as intelligent driving and video surveillance. For example, video target tracking can assist in planning the route of autonomous vehicles, detect targets in video surveillance, and predict future abnormal situations and possible safety hazards, providing theoretical guidance for safety prevention.

[0003] In the process of video target tracking, the traditional algorithm based on correlation filtering runs faster and has higher stability, but compared with a series of deep learning algorithms based on twin networks, the traditional algorithm is inferior in accuracy. Video target tracking is a major research hotspot in the field of computer vision. Although significant progress has been made in recent years, this research direction is still extremely challenging: 1) How to better deal with challenges such as target occlusion, drastic deformation or out of field of view, and solve the problem of contamination superposition caused by tracking models in video target tracking when the target is occluded, drastically deformed, or out of field of view; 2) How to effectively improve the discrimination ability of the twin network algorithm, that is, when the network judges similar targets, such as when multiple people or multiple vehicles appear in the scene, it often generates a situation in the feature layer where the confidence of similar targets is higher than that of the tracked target, causing the model to follow the wrong target, making it difficult to capture the target information and continue to track it, or the tracking fails due to misjudgment of similar targets; 3) In cluttered background scenes, the cluttered background often contaminates the tracking model, making it difficult to distinguish between the target and the background and effectively extract the target feature information, which leads to the loss of the target in subsequent frames. The related algorithms based on twin networks were first applied to tasks such as template matching and similarity measurement. In recent years, with the powerful performance of deep learning algorithms in the field of computer vision, algorithms based on deep learning twin networks have higher computing speeds than other deep learning algorithm frameworks. Therefore, it is of great value and significance to study target tracking algorithms based on deep learning twin networks. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art tracking model that when the target is obstructed, drastically deformed, or out of the field of view, pollution superposition will occur, similar targets will be difficult to distinguish, and cluttered backgrounds will easily cause contamination to the target. A target tracking method based on a deep learning twin network is provided, which aims to improve the tracking and positioning accuracy of the target in scenarios where target tracking is difficult, and to provide technical support for video target tracking.

[0005] To achieve the above objectives, the technical solutions provided by the present invention are as follows: A target tracking method based on a deep learning twin network includes the following steps: S1, obtain a video image file, specify a template target in the initial frame of the video image file, and give the true value of its target bounding box. The template target is used to provide a discriminant template for subsequent frames; S2, extract the feature information of the initial frame, process the feature information of the current frame through a deep separable convolutional network, obtain the corresponding local and global context information, and then perform feature enhancement to obtain the enhanced features of the template target ; S3, extracting the next subsequent frame as the current frame, extracting its feature information, processing the feature information of the current frame through the deep separable convolutional network to obtain the corresponding local and global context information, and then performing feature enhancement to obtain enhanced features of the subsequent frame ; S4, enhanced features of the subsequent frame through the twin network Enhanced features for template targets Perform association operation to obtain the strongly associated region image and its corresponding correlation response graph and loss function; S5, processing the strongly correlated region image through a classification regression network, iteratively outputting offset regression, and then obtaining the optimized correlation response map, target bounding box and corresponding loss function of the subsequent frame, and completing the target tracking of the current frame; S6, determine whether the current frame is the last frame, if not, return to step S3, if yes, output the target bounding box of all subsequent frames to complete video target tracking.

[0006] Further, in step S2 and step S3, the feature enhancement is specifically performed iteratively through m multi-scale large core modules, where m≥2; the multi-scale large core module includes a multi-scale large core attention module and a gated space attention module that are sequentially arranged, and the process of the multi-scale large core attention module is as follows:

[0007]

[0008] The process of the gated spatial attention module is as follows:

[0009]

[0010] in, is the feature obtained by the deep separable convolutional network, LN is the normalization operation, and N is the feature The normalized features, L is the feature The normalized features; are the learnable scaling factors in the current layer, is the outer product of the matrices, is the nth point-level convolution under the same dimension, n takes values ​​of 1, 2, 3, 4, 5, 6, which are recorded as , For dynamically adaptive large core attention, is the densely transformed feature and Features , is the output feature of the multi-scale large kernel attention module, It is the enhanced feature output by the multi-scale large kernel module.

[0011] Further, in step S2 and step S3, the Obtained by the following method: The feature N is processed by the first point-level convolution to obtain the feature , the fixed large-core attention processing is enhanced by the group multi-scale mechanism to obtain the feature The core attention , as shown below:

[0012] in, is the depthwise convolution, is a dilated convolution with a depth of D, It is point-level convolution; Features The i-th dimension feature The core attention Perform dynamic adaptation to obtain the dynamic features of the i-th dimension , as shown below:

[0013] in, for Space gate; Then the dynamic features Aggregate to get dynamically adapted large core attention .

[0014] Further, in step S2 and step S3, the Specifically obtained through the following methods: .

[0015] Furthermore, in step S2 and step S3, the deep separable convolutional network includes a feature decomposition module and a multi-order gated aggregation module; the feature information is input into the feature decomposition module, and the feature information is normalized, convolutional layered, and convolutional layered. , deep convolutional layer And the first activation function Processing, using different scaling rates to obtain low-order features , mid-level features and high-level features , specifically:

[0016]

[0017] Among them, X is the feature information, GAP is the global average pooling, For normalization, is a scaling factor initialized to zero, is a circular convolution; The multi-stage gated aggregation module is provided with an aggregation branch and a context branch. The aggregation branch is used to in Generate gating weights. The context branch is used to extract multi-scale features using circular convolutions of different sizes based on the generated gating weights, thereby capturing multi-order interactions in the context and converting low-order features into , mid-level features and high-level features As input feature X in Input multi-stage gated aggregation module, through the convolution layer And the first activation function Processing, while guided by the aggregation branch and the context branch, obtains feature Z as the output feature X out Output.

[0018] Furthermore, in step S5, the loss function is recorded as , obtained by the following formula:

[0019] in, is an indicator function, if If true, then the function value is 1, otherwise it is 0; For the loss of the classification branch, binomial cross entropy loss is used; is the actual position coordinate of the positive sample; Is it a positive sample? If it is a positive sample, it takes 1, and if it is a negative sample, it takes 0; For the loss of the regression branch, IOU loss is used; is the true value coordinate of the target position, is the normalized true value coordinate of the target position; is the total number of positive samples; is a modifiable parameter; is the predicted target location.

[0020] Further, in step S5, The value is 0.5.

[0021] The present invention also provides a computer program product, including a computer program, which is special in that when the program is executed by a processor, it implements the steps of the above-mentioned target tracking method based on a deep learning twin network.

[0022] At the same time, the present invention also provides a computer-readable storage medium on which a computer program is stored, the special feature of which is that when the program is executed by a processor, the steps of the above-mentioned target tracking method based on a deep learning twin network are implemented.

[0023] Beneficial effects of the present invention: 1. The target tracking method based on the deep learning twin network of the present invention sets up a multi-order deep separable convolutional network to capture local and global contextual information, effectively captures contextual information at different levels, and improves the limitations of traditional convolutional neural networks in context feature encoding.

[0024] 2. The target tracking method based on the deep learning twin network of the present invention sets a multi-scale large-core attention module and a gated spatial attention module to enhance the feature extraction capability in the target tracking method, so as to achieve better information aggregation and detail recovery.

[0025] 3. The target tracking method based on the deep learning twin network of the present invention obtains the relevant response map of subsequent frames through the classification regression network, and refines the position prediction of the target bounding box by outputting additional offset regression through the classification branch and the regression branch.

[0026] 4. The target tracking method based on deep learning twin network of the present invention adopts deep learning twin network, which makes target tracking accurate, efficient and flexible. It has better performance and algorithm robustness than traditional correlation filter algorithm, and can better cope with challenges such as target occlusion, drastic deformation, out of field of view, motion blur, etc., and provide technical support for video target tracking.

[0027] 5. The present invention also provides a computer program product and a computer-readable storage medium capable of executing the above method steps, which can promote and apply the method of the present invention to achieve target tracking on corresponding hardware devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flowchart of an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a feature decomposition module in an embodiment of the present invention; Figure 3 Schematic diagram of the structure of the multi-scale large core module in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] This method captures local and global contextual information by designing a multi-order deep separable convolutional network; then introduces a multi-scale large-core attention module and a gated spatial attention module to enhance the feature extraction capability in the image super-resolution task, so as to achieve better information aggregation and detail recovery; in the images of subsequent frames, a cross-correlation operation is performed between the template features and the features of the search area, and the strongly correlated area image is cropped; finally, the strongly correlated area image is input into a classification regression network, and the related response map in the subsequent frames is obtained through convolution operations. In conjunction with specific embodiments, the target tracking method based on a deep learning twin network of the present invention is described in detail below. Figure 1 The working process of the embodiment of the present invention is shown, which specifically includes the following steps: Step 1: Acquisition and preprocessing of video image files.

[0030] Video target tracking is performed by reading and identifying each frame of a video image file, which can be from a public data set or a non-public data set. This embodiment uses an existing public data set, selects one of the data sets such as OTB100, dtb70, VOT, and UAV123, and crops it. The data set includes multiple video image files, each of which corresponds to a video sequence. Multiple images in the same video sequence should have the same image resolution, and the image resolutions of different video sequences may be different. At the same time, a template target is specified in the image of the initial frame of each video sequence, and the true value of its target bounding box is given for feature extraction of the target in the initial frame. The template target provides a discriminant template for subsequent frames. The target bounding box includes the x-coordinates and y-coordinates of the four vertices of the rectangular box, where x represents the length direction of the image and y represents the height direction of the image.

[0031] In other embodiments of the present invention, for video target tracking using a non-public dataset, as with using a public dataset, each frame of the image is first cropped to provide the true value of the template target and its target bounding box in the initial frame.

[0032] Step 2: Extract the next frame of the video image file as the current frame, and extract feature information from the current frame. Specifically, a multi-order deep separable convolutional network is designed to capture local and global context information. The deep separable convolutional network consists of two modules: feature decomposition module and multi-order gated aggregation module. The details are as follows.

[0033] 2.1 Feature Decomposition Feature decomposition modules such as Figure 2 As shown, in order to make the network pay more attention to multi-order interactions, the feature information X of the current frame in each video image file is extracted, and the feature information X is input into the feature decomposition module, where , It is a real number matrix composed of three dimensions: length C, height H, and width W of the video image file. Through normalization (Norm), convolution layer , deep convolutional layer And the first activation function Processing, using different scaling rates to obtain low-order features , mid-level features and high-level features , thereby enhancing the expressiveness of features, where low-order features refer to fine-grained local textures in the image, mid-order features refer to complex global shapes, and high-order features refer to feature information X and low-order features through convolution operations. and mid-level features The processed features and detailed operations are as follows:

[0034]

[0035] Among them, GAP is the global average pooling, is the scaling factor initialized to zero, It is a circular convolution.

[0036] 2.2 Multi-stage gated aggregation The multi-stage gated aggregation module also has aggregation branches and context branches, such as Figure 2 As shown, the aggregation branch (Subtract) is used to in Generate gating weights. The context branch is used to extract multi-scale features using circular convolutions of different sizes based on the generated gating weights, thereby capturing multi-order interactions in the context and converting the low-order features output in step 2.1 into , mid-level features and high-level features As input features X in Input multi-stage gated aggregation module, through the convolution layer And the first activation function Processing, guided by the aggregation branch and the context branch, obtains feature Z as the output feature X out The output is as follows:

[0037]

[0038] in, is the second activation function, Concat is a circular convolution, which means that the low-order features , mid-level features and high-level features Merge, and then perform deep convolution to exchange information. It is the aggregated feature of low-order features, mid-order features and high-order features.

[0039] In the present invention, in order to adaptively aggregate the features extracted from the context branch, the outputs of the aggregation branch and the context branch are processed using the second activation function SiLU, which has both Sigmoid gating effect and stable training characteristics. The processing process is as follows:

[0040] Where u is the or ; Step 3: If Figure 3 As shown, feature enhancement is performed to improve the feature expression capability of the twin network. In order to better achieve information aggregation and detail recovery, the feature enhancement of the present invention performs iterative enhancement processing on the feature Z through m multi-scale large core modules to improve the feature expression capability in video target tracking. m ≥ 2 can be selected according to actual requirements for iterative enhancement processing until it converges to the pre-set algorithm accuracy parameters.

[0041] The multi-scale large core module includes a multi-scale large core attention module and a gated spatial attention module which are set in sequence. The specific processing process is as follows:

[0042] Among them, LN is the normalization operation, N is the feature The normalized features, L is the feature The normalized features are are respectively the learnable scaling factors in the current layer, is the outer product of the matrices, is the nth point-level convolution under the same dimension, n takes values ​​of 1, 2, 3, 4, 5, 6, which are recorded as , For dynamically adaptive large core attention, is the densely transformed feature and Features , is the output feature of the multi-scale large kernel attention module, It is the enhanced feature output by the multi-scale large kernel module.

[0043] Obtained by the following method: The attention mechanism adopted by the multi-scale large core attention module allows the twin network to focus on key information and ignore irrelevant information. The feature N is processed by the first point-level convolution to obtain the feature Input, in order to learn the attention map of full-scale information, the group multi-scale mechanism is used to enhance the fixed large kernel attention (LKA, Large kernel attention) to process it and obtain the feature The core attention , as shown below:

[0044] in, is the depthwise convolution, is a dilated convolution with a depth of D, It is a point-level convolution.

[0045] Features The i-th dimension feature Perform dynamic adaptation to obtain the dynamic features of the i-th dimension , and then the dynamic features Aggregate to get dynamically adapted large core attention . Specifically, the method is as follows: To avoid the block effect and learn more local information, the i-th dimension feature is transformed into The core attention Dynamic adaptation to dynamic features , as shown below:

[0046] in, for The space door, is the outer product of the matrices.

[0047] In order to obtain spatial information more effectively, the gated spatial attention module (GSAU) is used to first Perform a single-layer depth convolution to obtain the normalized feature L, and then perform the fourth and fifth point-level convolutions on the feature L to obtain the features and Features , and then densely transform the features and Features Weighted to get features , specifically expressed as:

[0048] By applying the gated spatial attention module, the nonlinear layer can be removed and local continuity can be captured while considering the complexity. After steps 2 and 3, the accuracy of feature extraction of the template object in the initial frame and the object in the subsequent frames can be effectively improved.

[0049] The enhanced features of the initial frame obtained through steps 2 and 3 are recorded as , and when the current frame is one of the subsequent frames, the enhanced features are recorded as ,After the subsequent frames obtain enhanced features through steps 2 and 3, they enter the next step.

[0050] Step 4: Perform association operations through the twin network to obtain a strongly associated region image: For the current frame, the main body of the twin network is used to enhance the features of the template target. Enhanced features with subsequent frames Share parameters between them, embed them into a common feature space, and correlate the features of the template area (i.e., the area within the target bounding box of the template target) and the features of the search area (i.e., the subsequent frame image) in the common feature space, judge their correlation, and obtain the strongly correlated area based on the correlation, crop the strongly correlated area to obtain the strongly correlated area image, and output the corresponding correlation response graph and loss function at the same time. The loss function is a loss function of similarity measurement. This is different from the pre-processed image cropping. The pre-processing is to crop the same video sequence into pictures of the same resolution, while the cropping here is to crop the areas with higher value of target feature similarity in subsequent frames after measuring the similarity of the features in subsequent frames to obtain strongly correlated area images. The strongly correlated area images will change arbitrarily in different frames according to the deformation and scale changes of the target. This is why the algorithm can be robust to changes in the size of the target. In this embodiment, the twin network is a deep learning twin network.

[0051] Video object tracking is considered as a similarity learning problem in the twin network framework. Specifically, the enhanced features of the template region u and the search region v are convolved and the corresponding features are obtained. and Features , and then establish the corresponding mapping function f(u,v) , then through the mapping function f(u,v) right and Perform similarity evaluation and obtain the relevant response map. Mapping function f(u,v) It can be expressed as:

[0052] in, Features and The similarity measurement function of .

[0053] when f(u,v) A higher value indicates that the targets on the template area and the search area are the same, while f(u,v) A lower value indicates that the targets in the template area and the search area are different. f(u,v) The area corresponding to the higher value is cropped and output to obtain a strongly correlated area image. In this embodiment, the size of the template area u is: 127×127×3, and the size of the search area v is 255×255×3. f(u,v) The value of is determined according to the loss function, and different judgment criteria can be used in different situations.

[0054] Step 5: Input the strongly correlated region image into the classification regression network. The strongly correlated region image is processed by the classification branch and regression branch in the classification regression network, and the offset regression is iteratively output. Then, the optimized correlation response map, target bounding box and corresponding loss function are obtained through the convolution operation to complete the target tracking of the current frame.

[0055] For each pixel in the strongly correlated region image, the classification branch classifies the corresponding image patch as a positive sample or a negative sample; the regression branch outputs an offset regression based on the positive and negative samples, and refines the position prediction of the target bounding box through the offset regression. The final optimized correlation response map size is 17×17×1. The classification regression network can use the existing classification regression network published in SiamFC++: Towards Robust and Accurate VisualTracking with Target Estimation Guidelines. (AAAI) published by Yinda Xu et al. in 2020.

[0056] The classification branch directly classifies the strongly correlated area image, and the regression branch regresses the target bounding box at the pixel position. In other words, each pixel position is directly regarded as a training sample. In this pixel-by-pixel prediction method, each pixel on the final feature map has only one prediction result, and each classification score directly represents the confidence of the target at the corresponding pixel.

[0057] In this embodiment, after obtaining the relevant response map in step 5, the classification score and the position regression score are obtained respectively, and the loss function corresponding to the target bounding box is recorded as , obtained by the following formula:

[0058] in, is an indicator function, if If true, then the function value is 1, otherwise it is 0; For the loss of the classification branch, binomial cross entropy loss is used. is the actual position coordinate of the positive sample, Is it a positive sample? If it is a positive sample, it takes 1, and if it is a negative sample, it takes 0; For the loss of the regression branch, IOU loss is used. is the true value coordinate of the target position, is the normalized true value coordinate of the target position, is the total number of positive samples, is a modifiable parameter, and in this embodiment, the value is 0.5. is the predicted target location.

[0059] Step 6: Determine whether the current frame is the last frame. If not, return to step 2. If yes, output the target bounding box of the template target corresponding to each frame in the subsequent frames, that is, the x and y coordinates of its four vertices, to complete the video target tracking.

[0060] To verify the effectiveness of the target tracking method based on deep learning twin network proposed in the present invention, this embodiment was trained on the GOT10K dataset and then tested on multiple other datasets, including OTB100, dtb70, VOT, UAV123 and other datasets. These datasets contain various tracking challenges, including scenes such as target occlusion, target deformation, out of field of view, motion blur, etc., and also cover image acquisition channels such as fixed cameras, mobile cameras, and drones. They all show good performance, achieved good accuracy results, and the fps can reach 80, meeting the real-time requirements of video target tracking.

[0061] The target tracking method based on the deep learning twin network of the present invention can also form a computer program product, which includes a computer program, and when the program is executed by a processor, it implements the steps of a target tracking method based on the deep learning twin network. At the same time, the target tracking method of the present invention can also be applied in a computer-readable storage medium, and the computer-readable storage medium stores a computer program. The above target tracking method can be stored in the computer-readable storage medium as a computer program, and when the computer program is executed by the processor, the steps of the above target tracking method are implemented.

Claims

1. A target tracking method based on deep learning twin network, characterized in that: The steps include: S1, obtain a video image file, specify a template target in the initial frame of the video image file, and give the true value of its target bounding box. The template target is used to provide a discriminant template for subsequent frames; S2, extract the feature information of the initial frame, process the feature information of the current frame through a deep separable convolutional network, obtain the corresponding local and global context information, and then perform feature enhancement to obtain the enhanced features of the template target ; S3, extract the next subsequent frame as the current frame and extract its feature information. The feature information of the current frame is processed by the deep separable convolutional network to obtain the corresponding local and global context information, and then the feature enhancement is performed to obtain the enhanced features of the subsequent frames. ; S4, enhanced features of the subsequent frame through the twin network Enhanced features for template targets Perform association operation to obtain the strongly associated region image and its corresponding correlation response graph and loss function; S5, processing the strongly correlated region image through a classification regression network, iteratively outputting offset regression, and then obtaining the optimized correlation response map, target bounding box and corresponding loss function of the subsequent frame, and completing the target tracking of the current frame; S6, determine whether the current frame is the last frame, if not, return to step S3, if yes, output the target bounding box of all subsequent frames to complete video target tracking.

2. According to claim 1, a target tracking method based on deep learning twin network is characterized in that: In step S2 and step S3, the feature enhancement is specifically performed through iterative enhancement processing by m multi-scale large core modules, where m≥2; the multi-scale large core module includes a multi-scale large core attention module and a gated space attention module that are sequentially arranged, and the process of the multi-scale large core attention module is as follows: ； ； The process of the gated spatial attention module is as follows: ； ； in, is the feature obtained by the deep separable convolutional network, LN is the normalization operation, and N is the feature The normalized features, L is the feature The normalized features; are the learnable scaling factors in the current layer, is the outer product of the matrices, is the nth point-level convolution under the same dimension, n takes values ​​of 1, 2, 3, 4, 5, 6, which are recorded as , For dynamically adaptive large core attention, is the densely transformed feature and Features , is the output feature of the multi-scale large kernel attention module, It is the enhanced feature output by the multi-scale large kernel module.

3. According to claim 2, a target tracking method based on deep learning twin network is characterized in that: In step S2 and step S3, the Obtained by the following method: The feature N is processed by the first point-level convolution to obtain the feature , the fixed large-core attention processing is enhanced by the group multi-scale mechanism to obtain the feature The core attention , as shown below: ； in, is the depthwise convolution, is a dilated convolution with a depth of D, It is point-level convolution; Features The i-th dimension feature The core attention Perform dynamic adaptation to obtain the dynamic features of the i-th dimension , as shown below: ； in, for Space gate; Then the dynamic features Aggregate to get dynamically adapted large core attention .

4. According to claim 3, a target tracking method based on deep learning twin network is characterized in that: In step S2 and step S3, the Specifically obtained through the following methods: 。 5. A target tracking method based on a deep learning twin network according to any one of claims 1 to 4, characterized in that: In step S2 and step S3, the deep separable convolutional network includes a feature decomposition module and a multi-order gated aggregation module; The feature information is input into the feature decomposition module, and then normalized and convolutional layers are used to generate the feature information. , deep convolutional layer And the first activation function Processing, using different scaling rates to obtain low-order features , mid-level features and high-level features , specifically: ； ； ； Among them, X is the feature information, GAP is the global average pooling, For normalization, is a scaling factor initialized to zero, is a circular convolution; The multi-stage gated aggregation module is provided with an aggregation branch and a context branch. The aggregation branch is used to in Generate gating weights. The context branch is used to extract multi-scale features using circular convolutions of different sizes based on the generated gating weights, thereby capturing multi-order interactions in the context and converting low-order features into , mid-level features and high-level features As input feature X in Input multi-stage gated aggregation module, through the convolution layer And the first activation function Processing, while guided by the aggregation branch and the context branch, obtains feature Z as the output feature X out Output.

6. According to claim 5, a target tracking method based on deep learning twin network is characterized in that: In step S5, the loss function is recorded as , obtained by the following formula: ； in, is an indicator function, if If true, then the function value is 1, otherwise it is 0; For the loss of the classification branch, binomial cross entropy loss is used; is the actual position coordinate of the positive sample; Is it a positive sample? If it is a positive sample, it takes 1, and if it is a negative sample, it takes 0; For the loss of the regression branch, IOU loss is used; is the true value coordinate of the target position, is the true value coordinate of the target position after normalization; is the total number of positive samples; is a modifiable parameter; is the predicted target location.

7. The target tracking method based on deep learning twin network according to claim 6 is characterized in that: In step S5, The value is 0.

5.

8. A computer program product, comprising a computer program, characterized in that: When the program is executed by a processor, the steps of a target tracking method based on a deep learning twin network as described in any one of claims 1 to 7 are implemented.

9. 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 a target tracking method based on a deep learning twin network as described in any one of claims 1 to 7 are implemented.

Images