Ground military target real-time tracking method based on DSST-Y

Abstract

The invention belongs to the technical field of computer vision. The invention provides a ground military target real-time tracking method based on DSST-Y. The ground military target real-time tracking method based on DSST-Y. According to the embodiment of the invention, the ground military target image is collected, and the construction of the ground military target data set is completed through image annotation. The original backbone network of the YOLOv5 model is replaced by the MobileNetV3 network to reduce the complexity of the network, the NAMAttention module is embedded in the neck of the network to enhance the feature extraction capability of the network, the SIoU loss function is adopted as a new bounding box regression loss function, and the model training speed and the network prediction accuracy are optimized. And the accuracy of the filter in the DSST algorithm is enhanced by introducing a reliability judgment mechanism. The DSST-Y tracking method provided by the invention can improve the tracking effect of the ground military target and avoid the drift phenomenon of the tracking target under the condition of ensuring the real-time performance, and has better robustness and reliability.

Description

Technical Field

[0001] The disclosed embodiments relate to the field of computer vision technology, specifically target detection and target tracking technology, and more particularly to a real-time tracking method for ground military targets based on DSST-Y. Background Art

[0002] Ground military target tracking is one of the important research contents in the field of target tracking. The core content is to obtain the category information and location information of enemy ground targets in real time from various backgrounds or environments. It is an important means to improve the battlefield information perception ability. Ground military targets have the characteristics of insufficient semantic information and high integration with the surrounding environment, which brings huge challenges to the target tracking task. Traditional target tracking methods mainly rely on manual features such as texture, color, shape, gradient, etc., and have problems such as low detection efficiency and poor robustness; the tracking method based on deep learning tracks the target through the model obtained by pre-training with a large number of samples. Although the accuracy is high, the real-time performance is difficult to meet the task requirements. Therefore, these two methods are difficult to be widely used in ground military target tracking in actual battlefields.

[0003] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0004] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the invention

[0005] The purpose of the embodiments of the present disclosure is to provide a real-time tracking method for ground military targets based on DSST-Y, thereby overcoming one or more problems caused by the limitations and defects of related technologies at least to a certain extent.

[0006] According to an embodiment of the present disclosure, a method for real-time tracking of ground military targets based on DSST-Y is provided, the method comprising: Collecting target image data, labeling the target image data, and dividing the labeled target image data into a training set, a validation set, and a test set according to a preset ratio; Build the YOLOv5-MNS model based on the MobileNetV3 network, NAMAttention module and SIoU loss function; The YOLOv5-MNS model is trained using the training set and the validation set to obtain the trained YOLOv5-MNS model, and the trained YOLOv5-MNS model is tested using the test set; Obtain a tracking image of the current frame, and use the trained YOLOv5-MNS model to detect and locate the target in the tracking image to obtain the category information and location information of the target; Based on the category information and the location information, the target is tracked using a DSST-Y tracking algorithm.

[0007] Furthermore, the steps of collecting target image data, labeling the target image data, and dividing the labeled target image data into a training set, a validation set, and a test set according to a preset ratio include: Selecting various types of ground military models, and collecting images corresponding to the ground military models in various environments to obtain target image data; Labeling the target image data to mark the category information and the location information of the target on each image; The labeled target image data is divided into the training set, the verification set and the test set according to a preset ratio.

[0008] Furthermore, the MobileNetV3 network includes: The first input layer, SE module and the first output layer; where The first input layer includes a 3x3 convolution layer for extracting features, and the h-swish activation function is used in the first input layer; The SE module includes several convolutional layers; The first output layer includes two 1x1 convolutional layers.

[0009] Furthermore, the NAM Attention module includes: Channel attention submodule and spatial attention submodule; among them, For the channel attention submodule, a normalized scaling factor is introduced, and the variance of the channel is measured using the scaling factor, as shown in the following formula:

[0010] in, is the variance of the channel, is the scale factor, represents the normalization operation, and Batch The mean and standard deviation of is the scale affine transformation parameter, is the displacement affine transformation parameter; and Batch The mean and standard deviation of is a constant; The characteristic expression of the channel attention submodule is:

[0011] in, is the feature of the channel attention submodule, Input feature map for channel attention, the first weight of each channel ; The characteristic expression of the spatial attention submodule is:

[0012] in, is the feature of the spatial attention submodule, Input feature map for spatial attention, is the scale factor, the second weight of the spatial pixel ; To suppress the weights, a regularization term is added to the loss function of the NAMAttention module as follows:

[0013] in, Indicates input, Indicates the output, is the network weight, is the loss function, is the penalty function, It's balance and punishment.

[0014] Furthermore, the SIoU loss function includes: Angle loss, distance loss, shape loss and IoU loss; among them, The expression of the angle loss is:

[0015] in, is the height difference between the center points of the real box and the predicted box, is the distance between the center points of the real box and the predicted box;

[0016]

[0017] in, is the real frame coordinate, is the predicted box coordinate; The expression of the distance loss is:

[0018] in, , , ; The expression of the shape loss is:

[0019] in, , , and are the width and height of the predicted box and the real box respectively, for the degree of concern about shape loss; The expression of the IoU loss is:

[0020] Among them, IoU represents IoU loss.

[0021] Furthermore, the method further comprises: Calculate the accuracy of the detection based on the category information, and determine whether the detection meets the standard based on the accuracy; If the target is met, the target is tracked; If the target is not met, the target is retested.

[0022] Further, based on the category information and the location information, the step of tracking the target using the DSST-Y tracking algorithm includes: For the tracking image of the current frame, initialize the DSST-Y tracking algorithm according to the category information and the position information thereof to obtain a filter; Obtaining a response value of the tracking image of a current frame according to the filter and the tracking image; Comparing the response value of the current frame with the average response value of the historical tracking results; If the response value is less than the response value mean, re-performing target detection on the tracking image of the current frame; If the response value is greater than or equal to the response value mean, then output and update the target position; It is determined whether the tracking image of the current frame is the tracking image of the last frame. If so, the target position is output; if not, target detection is performed on the tracking image of the next frame until tracking is completed.

[0023] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects: In the embodiment of the present disclosure, through the above-mentioned DSST-Y-based real-time tracking method for ground military targets, on the one hand, firstly, the ground military target image is collected, and the ground military target data set is constructed by image annotation. Then, the original backbone network of the YOLOv5 model is replaced by the MobileNetV3 network to reduce the complexity of the network, and the NAMAttention module is embedded in the Prediction module of the network to enhance the feature extraction capability of the network, and the SIoU loss function is used as the new bounding box regression loss function to optimize the model training speed and the accuracy of network prediction. Finally, the tracking object is automatically selected by YOLOv5-MNS to initialize the DSST-Y tracking algorithm, and a reliability judgment mechanism is introduced in each frame tracking. When the tracking result does not meet the requirements, the current frame image is detected by YOLOv5-MNS to determine a more accurate target position, and the DSST-Y tracking algorithm is reinitialized. The improved DSST-Y tracking algorithm can improve the tracking effect of ground military targets and avoid the occurrence of tracking target drift while ensuring real-time performance, and has good robustness and reliability. On the other hand, this method achieves accurate tracking of ground military targets in a variety of environments and effectively solves the problem of difficult detection of targets due to strong camouflage of ground military targets. The improved YOLOv5-MNS model is applied to ground military target detection, which improves the model's ability to meet real-time requirements while ensuring high accuracy when detecting ground military targets, which is conducive to expanding the application of deep learning algorithms in ground military target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0025] Figure 1 A step diagram showing a method for real-time tracking of ground military targets based on DSST-Y in an exemplary embodiment of the present disclosure; Figure 2 A structural diagram of a MobileNetV3 network in an exemplary embodiment of the present disclosure is shown; Figure 3 A structural diagram of a channel attention submodule in a NAM Attention module in an exemplary embodiment of the present disclosure is shown; Figure 4 A structural diagram showing a spatial attention submodule in a NAM Attention module in an exemplary embodiment of the present disclosure is shown; Figure 5 A structural diagram of a YOLOv5-MNS model in an exemplary embodiment of the present disclosure is shown; Figure 6 A flowchart for implementing a method for real-time tracking of ground military targets based on DSST-Y in an exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0026] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0027] In addition, the accompanying drawings are only schematic illustrations of the embodiments of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0028] This example implementation provides a real-time tracking method for ground military targets based on DSST-Y. Figure 1 As shown in , the real-time tracking method for ground military targets based on DSST-Y may include: steps S101 to S105.

[0029] Step S101: collecting target image data, labeling the target image data, and dividing the labeled target image data into a training set, a validation set, and a test set according to a preset ratio; Step S102: construct a YOLOv5-MNS model based on the MobileNetV3 network, NAMAttention module and SIoU loss function; Step S103: training the YOLOv5-MNS model using the training set and the validation set respectively to obtain the trained YOLOv5-MNS model, and testing the trained YOLOv5-MNS model using the test set; Step S104: obtaining a tracking image of the current frame, and using the trained YOLOv5-MNS model to detect and locate the target in the tracking image to obtain category information and position information of the target; Step S105: Based on the category information and the location information, the target is tracked using the DSST-Y tracking algorithm.

[0030] Through the above-mentioned real-time tracking method of ground military targets based on DSST-Y, on the one hand, firstly, the ground military target images are collected, and the ground military target dataset is constructed by image annotation. Then, the original backbone network of the YOLOv5 model is replaced by the MobileNetV3 network to reduce the complexity of the network, and the NAMAttention module is embedded in the Prediction module of the network to enhance the feature extraction ability of the network. The SIoU loss function is used as the new bounding box regression loss function to optimize the model training speed and the accuracy of network prediction. Finally, the tracking object is automatically selected by YOLOv5-MNS to initialize the DSST-Y tracking algorithm, and a reliability judgment mechanism is introduced in each frame tracking. When the tracking result does not meet the requirements, the current frame image is detected by YOLOv5-MNS to determine a more accurate target position, and the DSST-Y tracking algorithm is reinitialized. The improved DSST-Y tracking algorithm can improve the tracking effect of ground military targets and avoid the occurrence of tracking target drift while ensuring real-time performance, and has good robustness and reliability. On the other hand, this method achieves accurate tracking of ground military targets in a variety of environments and effectively solves the problem of difficult detection of targets due to strong camouflage of ground military targets. The improved YOLOv5-MNS model is applied to ground military target detection, which improves the model's ability to meet real-time requirements while ensuring high accuracy when detecting ground military targets, which is conducive to expanding the application of deep learning algorithms in ground military target detection.

[0031] Next, we will refer to Figures 1 to 6 Each step of the above-mentioned DSST-Y-based real-time tracking method for ground military targets in this example implementation is described in more detail.

[0032] In step S101, a ground military target dataset is constructed. YOLOv5, a target detection algorithm based on deep learning, belongs to supervised learning, so a large amount of data is required as support during model training. The dataset used in this application is collected for a specific purpose, and the specific steps are as follows: First, four common ground military models (tanks, armored vehicles, missile launchers, and self-propelled artillery) are selected to collect target data in roads, camouflage nets, grass, bushes, and different weather environments. The data type is pictures. The model trained with these data is robust and has strong generalization capabilities.

[0033] Secondly, label the data: use Labelimg software to label the collected ground military target data and generate a label file in the form of a txt file with the target location information and category information on each picture.

[0034] Finally, data set division: the processed ground military target data set is divided into training set, validation set and test set in a ratio of 8:1:1 for subsequent model training and model evaluation.

[0035] The construction of the ground military target dataset is completed through step S101 and is used for the subsequent training of the YOLOv5 target detection model.

[0036] In step S102, MobileNetV3 is used to replace the original backbone network in the YOLOv5 model. Although YOLOv5 has good detection accuracy, especially when it is used on mobile platforms, due to its efficient computing power and the large YOLOv5 target detection network model, it does not meet the real-time requirements for target detection in practical applications.

[0037] MobileNetV3 is optimized based on the advantages of MobileNetV1 and MobileNetV2. Its structure is as follows Figure 2 As shown in the figure. MobileNetV3 introduces the channel attention model SE, which further optimizes the feature representation by learning a set of weights and assigning weights to each channel, so that important features are strengthened and other features are weakened. It also redesigns the activation function, that is, it uses the Hardswish activation function with lower computational complexity than the original swish activation function. Therefore, while maintaining a high accuracy, the network further reduces the size and computational complexity of the model, and has better real-time detection performance.

[0038] Introducing the NAMAttention mechanism. This application adopts an efficient and lightweight attention mechanism - NAMAttention, whose structure is as follows Figure 3 and Figure 4 As shown. Among them, Figure 3 This is the structural diagram of the channel attention submodule in the NAM Attention module; Figure 4This is the structural diagram of the spatial attention submodule in the NAMAttention module. This attention mechanism uses the module integration of CBAM and redesigns the spatial and channel attention submodules. The spatial attention allows the neural network to pay more attention to the pixel areas in the image that are effective for classification and ignore other insignificant areas. The channel attention is used to process the distribution relationship of the feature map channels. This attention mechanism performs attention feature fusion on the given feature map from the two dimensions of channel and space, thus saving parameters and computing power to the greatest extent and being easy to integrate into other network structures. Under the action of the NAMAttention attention mechanism, the new feature map will obtain attention weights in two different dimensions, channel and space, greatly improving the connection between each feature in channel and space, which is more conducive to extracting effective features of the target, thereby improving the accuracy of network target detection.

[0039] In summary, the NAMAttention mechanism has the following three advantages: 1. The NAM Attention module imposes a weight sparsity penalty on the attention mechanism by adding a regularization term to the loss function to suppress less significant weights, thereby improving the generalization ability and efficiency of the model.

[0040] 2. The mechanism uses the scaling factor of batch normalization to weigh the importance of channels and pixels, avoiding the fully connected and convolutional layers used in attention mechanisms such as SE, BAM, and CBAM, thereby improving computational efficiency.

[0041] 3. The NAM Attention module combines the channel attention and spatial attention sub-modules, and uses the scaling factor of batch normalization to weigh the importance of channels and pixels respectively.

[0042] In the NAMAttention mechanism, for the channel attention submodule, a scaling factor of normalization (BN) is used, which measures the variance of channels and represents their importance, as shown in the following formula:

[0043] in, and They are The mean and standard deviation of and are the trainable affine transformation parameters (scale and shift).

[0044] Then, the characteristic expression of the channel attention submodule is:

[0045] Among them, the weight .

[0046] Correspondingly, the expression of the spatial attention submodule is:

[0047] in, is the scale factor, weight .

[0048] In order to suppress less significant weights, a regularization term is added to the loss function as follows:

[0049] in, Indicates input, Indicates the output, is the network weight, is the loss function, is the penalty function, It's balance and punishment.

[0050] Replace the loss function. The SIoU loss function takes into account the center point deviation and shape deviation while calculating the IoU, and further introduces the factor of direction perception, so that it can better handle non-square aspect ratios, thereby improving accuracy and robustness in object detection. These extensions can more accurately reflect the difference between the predicted box and the real box, especially in complex backgrounds and object detection scenarios of different scales.

[0051] The SIoU loss function consists of four parts: angle loss, distance loss, shape loss, and IoU loss.

[0052] 1. Angle loss

[0053] in, is the height difference between the center points of the real box and the predicted box, is the distance between the two center points.

[0054]

[0055]

[0056] in, is the real frame coordinate, is the predicted box coordinate.

[0057] 2. Distance loss

[0058] in, , , .

[0059] 3. Shape loss

[0060] in, , , and are the width and height of the predicted box and the real box respectively, is the degree of concern about shape loss.

[0061] 4. IoU Loss The full name of IoU is Intersection over Union, which is the ratio of the intersection and union of the "predicted bounding box" and the "real bounding box". The calculation formula of IoU is:

[0062] It can be concluded that the expression of SIoU loss function is defined as follows:

[0063] In step S103, the multi-scale training sets several images of different scales, and randomly selects one scale for training every certain iterations. The trained model has good generalization ability and strong robustness.

[0064] At the beginning of model training, we first customize a scale list scale_list = [168,192,224,256], which indicates the possible sizes to which the image may be adjusted during training. In each iteration, the selection function randomly determines a size and adjusts the size of the training image accordingly. This measure can make the model adapt to inputs of different sizes, thereby improving its generalization ability and robustness.

[0065] The YOLOv5-MNS network structure obtained by the above improvement strategy is as follows Figure 5 shown.

[0066] In step S104 and step S105, the conventional target detection task only needs (Average precision mean) and (number of detection frames per second) can meet the requirements. However, for ground military targets with strong camouflage, its (missed alarm) and The (false alarm) index has a great influence on the battlefield situation judgment, so these two types of indicators need to be used as one of the evaluation indicators. The calculation formula is as follows:

[0067]

[0068]

[0069]

[0070] in It is The number of camouflaged targets contained in the secondary image, For the detection algorithm The number of correctly predicted ground military targets in the secondary image, The detection algorithm will The number of ground military targets identified in the background of the secondary image, is the total number of images involved in the detection.

[0071] It can be seen from Table 1 that the YOLOv5-MNS algorithm improved by the experiment is better than the original algorithm The value increased by 6.16%, and the missed detection rate ( ) reduced by 4.9%, false alarm rate ( ) is reduced by 0.46%, and still has better overall performance compared with other detection algorithms.

[0072] Table 1 Comparison results before and after network improvement

[0073] DSST algorithm initialization: At the beginning of tracking, the YOLOv5 target detection model is used to identify and locate the tracking target. When the confidence level is not less than 0.9 (the maximum is 1), the detection result is considered accurate. At this time, the obtained category information and location information are used to initialize the DSST tracking algorithm to obtain an accurate filter. Target tracking: Use the filter to perform correlation operations with the image on the next video frame. The area with the largest response value is the location of the target on the current frame, and its response value is saved at the same time. .

[0074] Tracking result judgment: then compare and The size of , use YOLOv5 to detect the current frame image to determine the exact location of the target and reinitialize KCF, otherwise it will be trained normally and the filter will be updated.

[0075] Figure 6 This is a flow chart of a method for real-time tracking of ground military targets based on DSST-Y provided in the implementation example of this application.

[0076] This application replaces the original backbone structure of the YOLOv5 network, adds the NAMAttention mechanism, and finally uses SIoU as a new loss function, which effectively solves the problem of poor real-time performance of the network in ground military target detection, and its detection accuracy is also improved; at the same time, this application proposes a ground military target detection method based on DSST-Y, which automatically selects the tracking object through YOLOv5-MNS to initialize the DSST tracking algorithm, and introduces a reliability judgment mechanism in each frame tracking. When the tracking result does not meet the requirements, YOLOv5-MNS is used to detect the current frame image to determine a more accurate target position, and the DSST algorithm is reinitialized. The improved DSST-Y tracking method can improve the tracking effect of ground military targets and avoid the occurrence of tracking target drift while ensuring real-time performance, and has good robustness and reliability.

[0077] Through the above-mentioned real-time tracking method of ground military targets based on DSST-Y, on the one hand, firstly, the ground military target images are collected, and the ground military target dataset is constructed by image annotation. Then, the original backbone network of the YOLOv5 model is replaced by the MobileNetV3 network to reduce the complexity of the network, and the NAMAttention module is embedded in the Prediction module of the network to enhance the feature extraction ability of the network. The SIoU loss function is used as the new bounding box regression loss function to optimize the model training speed and the accuracy of network prediction. Finally, the tracking object is automatically selected by YOLOv5-MNS to initialize the DSST-Y tracking algorithm, and a reliability judgment mechanism is introduced in each frame tracking. When the tracking result does not meet the requirements, the current frame image is detected by YOLOv5-MNS to determine a more accurate target position, and the DSST-Y tracking algorithm is reinitialized. The improved DSST-Y tracking algorithm can improve the tracking effect of ground military targets and avoid the occurrence of tracking target drift while ensuring real-time performance, and has good robustness and reliability. On the other hand, this method achieves accurate tracking of ground military targets in a variety of environments and effectively solves the problem of difficult detection of targets due to strong camouflage of ground military targets. The improved YOLOv5-MNS model is applied to ground military target detection, which improves the model's ability to meet real-time requirements while ensuring high accuracy when detecting ground military targets, which is conducive to expanding the application of deep learning algorithms in ground military target detection.

[0078] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0079] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0080] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A real-time tracking method for ground military targets based on DSST-Y, characterized in that: The method includes: Collecting target image data, labeling the target image data, and dividing the labeled target image data into a training set, a validation set, and a test set according to a preset ratio; Build the YOLOv5-MNS model based on the MobileNetV3 network, NAMAttention module and SIoU loss function; The YOLOv5-MNS model is trained using the training set and the validation set to obtain the trained YOLOv5-MNS model, and the trained YOLOv5-MNS model is tested using the test set; Obtain a tracking image of the current frame, and use the trained YOLOv5-MNS model to detect and locate the target in the tracking image to obtain the category information and location information of the target; Based on the category information and the location information, the target is tracked using a DSST-Y tracking algorithm.

2. The real-time tracking method for ground military targets based on DSST-Y according to claim 1 is characterized in that: The steps of collecting target image data, labeling the target image data, and dividing the labeled target image data into a training set, a validation set, and a test set according to a preset ratio include: Selecting various types of ground military models, and collecting images corresponding to the ground military models in various environments to obtain target image data; Labeling the target image data to mark the category information and the location information of the target on each image; The labeled target image data is divided into the training set, the verification set and the test set according to a preset ratio.

3. The real-time tracking method for ground military targets based on DSST-Y according to claim 1 is characterized in that: The MobileNetV3 network includes: The first input layer, SE module and the first output layer; where The first input layer includes a 3x3 convolution layer for extracting features, and the h-swish activation function is used in the first input layer; The SE module includes several convolutional layers; The first output layer includes two 1x1 convolutional layers.

4. The real-time tracking method for ground military targets based on DSST-Y according to claim 3 is characterized in that: The NAMAttention module includes: Channel attention submodule and spatial attention submodule; among them, For the channel attention submodule, a normalized scaling factor is introduced, and the variance of the channel is measured using the scaling factor, as shown in the following formula: in, is the variance of the channel, is the scale factor, represents the normalization operation, and Batch The mean and standard deviation of is the scale affine transformation parameter, is the displacement affine transformation parameter; and Batch The mean and standard deviation of is a constant; The characteristic expression of the channel attention submodule is: in, is the feature of the channel attention submodule, Input feature map for channel attention, the first weight of each channel ; The characteristic expression of the spatial attention submodule is: in, is the feature of the spatial attention submodule, Input feature map for spatial attention, is the scale factor, the second weight of the spatial pixel ; To suppress the weights, a regularization term is added to the loss function of the NAMAttention module as follows: in, Indicates input, Indicates the output, is the network weight, is the loss function, is the penalty function, It's balance and punishment.

5. The real-time tracking method for ground military targets based on DSST-Y according to claim 4 is characterized in that: The SIoU loss function includes: Angle loss, distance loss, shape loss and IoU loss; among them, The expression of the angle loss is: in, is the height difference between the center points of the real box and the predicted box, is the distance between the center points of the real box and the predicted box; in, is the real frame coordinate, is the predicted box coordinate; The expression of the distance loss is: in, , , ; The expression of the shape loss is: in, , , and are the width and height of the predicted box and the real box respectively, for the degree of concern about shape loss; The expression of the IoU loss is: Among them, IoU represents IoU loss.

6. The real-time tracking method for ground military targets based on DSST-Y according to claim 1 is characterized in that: The method further includes: Calculate the accuracy of the detection based on the category information, and determine whether the detection meets the standard based on the accuracy; If the target is met, the target is tracked; If the target is not met, the target is retested.

7. The real-time tracking method for ground military targets based on DSST-Y according to claim 1 is characterized in that: The step of tracking the target using the DSST-Y tracking algorithm based on the category information and the location information includes: For the tracking image of the current frame, initialize the DSST-Y tracking algorithm according to the category information and the position information thereof to obtain a filter; Obtaining a response value of the tracking image of a current frame according to the filter and the tracking image; Comparing the response value of the current frame with the average response value of the historical tracking results; If the response value is less than the response value mean, re-performing target detection on the tracking image of the current frame; If the response value is greater than or equal to the response value mean, then output and update the target position; It is determined whether the tracking image of the current frame is the tracking image of the last frame. If so, the target position is output; if not, target detection is performed on the tracking image of the next frame until tracking is completed.

Images