Unmanned aerial vehicle target tracking method based on space-time interaction and long and short time attention

Abstract

The invention discloses an unmanned aerial vehicle target tracking method based on space-time interaction and long and short time attention, and the method comprises the steps: constructing a tracking model, tracking a target through the model, and updating a target template, and the construction of the tracking model comprises the construction of a backbone network, the feature mapping output from the backbone network, and the flattening processing of a bottleneck structure before the input of the backbone network. The feature sequence is input into an LSDA module, global and local information of the feature sequence is obtained, a decoder obtains a target from an encoder to inquire an enhanced feature sequence as input, and the input is sent to a frame prediction head to predict object frame coordinates. According to the method, self-attention in an original Transform model is replaced, the target tracking performance is improved, meanwhile, the model calculation amount is reduced, and the method is more suitable for unmanned aerial vehicle target tracking with the complex view angle and limited calculation resources.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle tracking, and in particular to a method for tracking unmanned aerial vehicle targets based on spatiotemporal interaction and long- and short-term attention. Background Art

[0002] Visual target tracking is to predict the position and size of a target in subsequent frames based on the initial state of a given target. It is an important research direction in computer vision and has important applications in human-computer interaction, intelligent video surveillance, etc. With the widespread use of drone technology in military and civilian fields, the demand for intelligent drone tracking systems with high accuracy, low power consumption and good real-time performance is also increasing. Currently, most popular trackers are based on a deep Siamese network architecture, which uses correlation operations to consider the similarity between the template branch and the search branch. However, the correlation operation itself is a linear matching process, which leads to the loss of semantic information, which limits the tracker from capturing the complex nonlinear interactions between the template and the search area. As a neural network model based on the self-attention mechanism, Transformer can well model long-distance dependencies. It replaces related operations with self-attention to achieve feature fusion and enhancement. It is widely used in the field of target tracking and effectively improves the performance of tracking algorithms. Due to the unique aerial perspective of drones and complex and changeable application scenarios, the current Transformer-based tracking algorithm is still applied to the field of drone target tracking. There are still the following major problems.

[0003] First, the tracking algorithm based on the Transformer model mainly uses the attention mechanism to enhance and fuse features to improve performance, but it is insufficient in extracting local detail features. For example, the CGCAD and SiamAttn algorithms introduce relevant guided attention and self-attention for identification tracking, and TransT and TrDiMP track by enhancing the target template and search area features. Due to the lack of attention to relevant local information in the target area, it is easy to lose the target when there is a complex background or similar target interference. Secondly, the Transformer-based tracker usually only considers the spatial appearance characteristics of the target and the search area, ignoring the extraction and utilization of time domain information. Under the complex and changeable tracking perspective of the drone, the robustness of the target model is insufficient. Finally, the attention layer in the Transformer model models long-distance dependencies by performing dot product operations on all input elements, which is very computationally intensive and makes it difficult to deploy the model on a drone mobile platform with limited computing power.

[0004] In view of the above problems, the present invention provides a drone target tracking method based on spatiotemporal interaction and long- and short-term attention, which can adapt to the requirements of drone target tracking scenarios. Summary of the invention

[0005] The present invention provides a UAV target tracking method based on spatiotemporal interaction and long- and short-term attention, which can adapt to the requirements of UAV target tracking scenarios.

[0006] The purpose of the present invention is to provide a method for tracking a target of a drone based on spatiotemporal interaction and long- and short-term attention, including building a tracking model, model tracking of a target and updating a target template. The tracking model building includes building a backbone network, and performing a "flattening" process on the bottleneck structure before inputting the feature map output from the backbone network. The feature sequence is input into the LSDA module to obtain the global and local information of the feature sequence. The decoder obtains a target query and an enhanced feature sequence from the encoder as input, and the input is sent to a frame prediction head to predict the coordinates of the object frame. The model tracking of the target includes framing the location of the target object, and then fine-tuning the regression frame, so as to train the model to predict the target position and add the intersection and union ratio of key point distances as a frame regression function to the TLSDA algorithm. The target template updating includes designing a target state perception module and updating the target template.

[0007] Furthermore, the specific steps for building the backbone network are as follows: Step 1: The input of the backbone network is a triplet of images: the template image of the initial target object , Search area of ​​the current frame and dynamic templates ; Step 2: After passing the main trunk, the template z , Search images x and dynamic templates t are mapped to , and , where the initial template provides spatial information, and the dynamic template can capture the changes in target appearance over time, thereby providing additional temporal information.

[0008] Furthermore, the bottleneck structure is “flattened” into a “flat” structure, where attention occupies the main computation in the “flat” Transformer model without changing the channel dimension, achieving comparable performance under large computational load, which can be reduced through further optimization.

[0009] Furthermore, the specific steps of the LSDA module are as follows: Step 1: Capture global and local contexts separately. The LSDA module follows a two-branch design, where the left branch captures the global context, while the right branch models the local context. Step 2: Split the features into two parts along the channel dimension and mix them by the following FFN layer, which reduces the overall computational effort by half. Step 3: The left branch is a normal attention module, but the channel dimension is reduced by half. For the right branch that captures local relations, a natural idea is to apply convolution on the sequence. Step 4: By sliding the window, the diagonal line is covered; Step 5: To further reduce the amount of computation, replace the ordinary convolution with a lighter version consisting of a linear layer and a depth-wise convolution. Step 6: By placing the attention and convolution modules side by side, the model is encouraged to process the feature sequence from two different perspectives: global and local, so that the architecture can benefit from focus.

[0010] Furthermore, the decoder is different from DETR, which uses 100 object queries. The model only inputs one query in the decoder to predict a bounding box of the target object. There is only one prediction, and the Hungarian algorithm used to predict associations in DETR is deleted. The decoder stacks N decoder layers, each of which consists of a self-attention parallel convolution, an encoder-decoder attention, and a feedforward network. In the Encoder-Decoder attention module, the target query can focus on all positions on the template and the search area features, thereby learning a robust representation of the final bounding box prediction.

[0011] Furthermore, the specific steps of the bounding box prediction head are as follows: Step 1: Extract search area features from the output sequence of the encoder; Step 2: Calculate the similarity between the search region features and the decoder’s output embedding; Step 3: Multiply the similarity score by the feature element of the search area to enhance the important areas and weaken the areas with poor discriminability, and reshape the new feature sequence into The feature map of Step 4: Input a simple fully convolutional network FCN, which consists of L stacked Conv-BN-ReLU layers, and outputs two probability maps for the upper left corner and lower right corner of the object bounding box respectively. and ,By estimating the probability distribution of border corner points, the ,quality of border estimation is improved.

[0012] Furthermore, the specific calculation steps of the key point distance intersection are as follows: Step 1: Frame and frame They are the real box and the predicted box respectively, and the line segments between the corresponding vertices of the two boxes are recorded , , The length is , , , remember the line segment between the center points of the two frames The length is , respectively extend the length of the two frames to obtain the minimum covering frame, and connect the points and Point , get the diagonal of the minimum covering box , the length is , The calculation is done by Adding a penalty term on the basis of: (1) in, , to ensure Non-negativity of values, penalty term It must be a value between [0, 1]; Step 2: Calculate the sum of the squares of the Euclidean distances between four groups of points, including three pairs of vertices and one pair of center points. Then, normalize the sum to ensure the normalization of the penalty term. The normalized benchmark meets two conditions: (1) The normalized benchmark is used as the denominator to compare with the sum result, so the benchmark cannot be 0; (2) To ensure that the penalty term is a value distributed between [0, 1], the normalized benchmark must be larger than the value of the above sum result. The diagonal length 𝑐 of the minimum coverage box between the predicted box and the true box is selected as the normalized benchmark. The minimum value of 𝑐 is the diagonal length of the true box, which meets condition 1; in addition, the diagonal length 𝑐 of the minimum coverage box is the longest distance between any two points in the two boxes, which is greater than the distance between the three vertices and the center point corresponding to the two boxes, which meets condition 2. In addition, the use of the Euclidean square of the distance between points is larger than the direct use of the distance between points, which improves the calculation sensitivity of the penalty term. The calculation formula of the penalty term is: (2) in, , , , , Respectively represent , , , and The square of the Euclidean distance between the two boxes, the numerator is the Euclidean distance between the four groups of corresponding key points of the two boxes, obtained by formula (1) and formula (2): The calculation formula is: (3) Step 3: From formula (3), we can get: (1) ,when The value is 1, and the penalty value is 0, that is, when the two frames overlap, The value is 1; when The value is 0, and the penalty value is 1, that is, when the two boxes have no intersection, The value is -1, (2) due to It is a range value, not sensitive to scale, and will not change due to image scale issues. Value (3) When the two boxes do not intersect, Using the distance between points as gradient optimization makes up for The shortcomings, The loss function is: (4) Step 4: The loss function uses the ideal intersection-union ratio minus the actual calculated intersection-union ratio. This difference is calculated as the loss. Loss function, when the two boxes overlap, If the value is 1, The value is 0; when the two boxes have no intersection, The value of is -1, The value is 2, so , the predicted box and the true box The closer it is to 0, the closer the morphological distance between the two frames is, otherwise the distance and similarity between the two frames are lower.

[0013] Furthermore, the specific design steps of the loss function are as follows: Step 1: The tracker is trained in an end-to-end manner, combining L1 loss and KIoU loss, and the loss function is: (5) in, and Represent the basic value and prediction box respectively, , is a hyperparameter; Step 2: Joint learning of localization and classification may lead to suboptimal solutions for both tasks. This helps to decouple localization and classification. Therefore, the training process is divided into two stages, with localization as the main task and classification as the secondary task. In the first stage, except for the scoring head, the entire network only trains the localization-related loss in equation (5). In this stage, all search images contain the target object and the model learns the localization ability. In the second stage, only the scoring head is optimized, and the binary cross entropy loss is defined as: (6) in, For the base tag, is the prediction confidence.

[0014] Furthermore, the specific steps for updating the target template are as follows: Step 1: During the tracking process, in each frame, the target template is cropped from the current image and then input into the backbone for feature extraction. The network takes a search area from the current frame as input, determines whether to update the target template through the target perception module, and returns the predicted box as the final result without using any post-processing such as cosine window or bounding box smoothing. Step 2: By effectively utilizing the target's spatiotemporal context interaction information, the target template is dynamically updated to avoid tracking drift caused by complex environmental factors in UAV application scenarios and further improve the robustness of tracking.

[0015] The present invention has the following advantages: the present invention designs a "flat" structure, further improves the model's ability to capture contextual relationships through spatiotemporal interaction, and optimizes the design of the bottleneck structure.

[0016] The present invention replaces the self-attention in the original Transformer model, improves the target tracking performance while reducing the model calculation amount, and is more suitable for UAV target tracking with complex viewing angles and limited computing resources.

[0017] The present invention uses an intersection-over-union (IRU) box regression method based on key point distance to improve the positioning accuracy in target tracking. Experimental results on multiple challenging data sets show that the amount of computation can be reduced and the tracking performance can be improved.

[0018] The present invention designs a dynamic template update mechanism, introduces a target positioning reliability evaluation module, and determines whether the current frame template is updated according to the reliability scores of target positioning of several consecutive frames. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a tracking framework diagram of the present invention; Figure 2 It is a frame prediction structure diagram of the present invention; Figure 3 This is a "flat" structural diagram of the present invention; Figure 4 It is the calculation amount diagram of the present invention; Figure 5 For the general attention of the present invention; Figure 6 is the attention graph in the LSDA of the present invention; Figure 7 The block structure diagram of the lightweight Transformer of the present invention; Figure 8 This is a schematic diagram of the key point distance intersection and union ratio calculation principle of the present invention. DETAILED DESCRIPTION

[0020] The present invention provides a method for tracking a target of an unmanned aerial vehicle based on spatiotemporal interaction and long- and short-term attention, including building a tracking model, model tracking of a target and updating a target template. The tracking model building includes building a backbone network, performing a bottleneck structure "flattening" processing on a feature map output from the backbone network before inputting the backbone network, inputting a feature sequence into an LSDA module, obtaining global and local information of the feature sequence, a decoder obtaining a target query and an enhanced feature sequence from an encoder as input, and sending the input to a frame prediction head to predict the coordinates of an object frame, model tracking of a target includes framing the location of a target object, and then fine-tuning a regression frame, thereby training a model to predict a target position and adding a key point distance intersection-union ratio as a frame regression function into a TLSDA algorithm, and updating a target template includes designing a target state perception module and updating a target template.

[0021] In this embodiment, the specific steps of building the backbone network are as follows: Step 1: The input of the backbone network is a triplet of images: the template image of the initial target object , Search area of ​​the current frame and dynamic templates ; Step 2: After passing the main trunk, the template z , Search images x and dynamic templates t are mapped to , and , where the initial template provides spatial information, and the dynamic template can capture the changes in target appearance over time, thereby providing additional temporal information.

[0022] Furthermore, the bottleneck structure is “flattened” into a “flat” structure, where attention occupies the main computation in the “flat” Transformer model without changing the channel dimension, achieving comparable performance under large computational load, which can be reduced through further optimization.

[0023] In this embodiment, the specific steps of the LSDA module are as follows: Step 1: Capture global and local contexts separately. The LSDA module follows a two-branch design, where the left branch captures the global context, while the right branch models the local context. Step 2: Split the features into two parts along the channel dimension and mix them by the following FFN layer, which reduces the overall computational effort by half. Step 3: The left branch is a normal attention module, but the channel dimension is reduced by half. For the right branch that captures local relations, a natural idea is to apply convolution on the sequence. Step 4: By sliding the window, the diagonal line is covered; Step 5: To further reduce the amount of computation, replace the ordinary convolution with a lighter version consisting of a linear layer and a depth-wise convolution. Step 6: By placing the attention and convolution modules side by side, the model is encouraged to process the feature sequence from two different perspectives: global and local, so that the architecture can benefit from focus.

[0024] In this embodiment, the decoder is different from DETR which uses 100 object queries. The model only inputs one query in the decoder to predict a bounding box of the target object. There is only one prediction. The Hungarian algorithm used to predict associations in DETR is deleted. The decoder stacks N decoder layers, each of which consists of a self-attention parallel convolution, an encoder-decoder attention, and a feedforward network. In the Encoder-Decoder attention module, the target query can focus on all positions on the template and the search area features, thereby learning a robust representation of the final bounding box prediction.

[0025] In this embodiment, the specific steps of the border prediction head are as follows: Step 1: Extract search area features from the output sequence of the encoder; Step 2: Calculate the similarity between the search region features and the decoder’s output embedding; Step 3: Multiply the similarity score by the feature element of the search area to enhance the important areas and weaken the areas with poor discriminability, and reshape the new feature sequence into The feature map of Step 4: Input a simple fully convolutional network FCN, which consists of L stacked Conv-BN-ReLU layers, and outputs two probability maps for the upper left corner and lower right corner of the object bounding box respectively. and ,By estimating the probability distribution of border corner points, the ,quality of border estimation is improved.

[0026] In this embodiment, the specific steps for calculating the intersection of key point distances are as follows: Step 1: Frame and frame They are the real box and the predicted box respectively, and the line segments between the corresponding vertices of the two boxes are recorded , , The length is , , , remember the line segment between the center points of the two frames The length is , respectively extend the length of the two frames to obtain the minimum covering frame, and connect the points and Point , get the diagonal of the minimum covering box , the length is , The calculation is done by Adding a penalty term on the basis of: (1) in, , to ensure Non-negativity of values, penalty term It must be a value between [0, 1]; Step 2: Calculate the sum of the squares of the Euclidean distances between four groups of points, including three pairs of vertices and one pair of center points. Then, normalize the sum to ensure the normalization of the penalty term. The normalized benchmark meets two conditions: (1) The normalized benchmark is used as the denominator to compare with the sum result, so the benchmark cannot be 0; (2) To ensure that the penalty term is a value distributed between [0, 1], the normalized benchmark must be larger than the value of the above sum result. The diagonal length 𝑐 of the minimum coverage box between the predicted box and the true box is selected as the normalized benchmark. The minimum value of 𝑐 is the diagonal length of the true box, which meets condition 1; in addition, the diagonal length 𝑐 of the minimum coverage box is the longest distance between any two points in the two boxes, which is greater than the distance between the three vertices and the center point corresponding to the two boxes, which meets condition 2. In addition, the use of the Euclidean square of the distance between points is larger than the direct use of the distance between points, which improves the calculation sensitivity of the penalty term. The calculation formula of the penalty term is: (2) in, , , , , Respectively represent , , , and The square of the Euclidean distance between the two boxes, the numerator is the Euclidean distance between the four groups of corresponding key points of the two boxes, obtained by formula (1) and formula (2): The calculation formula is: (3) Step 3: From formula (3), we can get: (1) ,when The value is 1, and the penalty value is 0, that is, when the two frames overlap, The value is 1; when The value is 0, and the penalty value is 1, that is, when the two boxes have no intersection, The value is -1, (2) due to It is a range value, not sensitive to scale, and will not change due to image scale issues. Value (3) When the two boxes do not intersect, Using the distance between points as gradient optimization makes up for The shortcomings, The loss function is: (4) Step 4: The loss function uses the ideal intersection-union ratio minus the actual calculated intersection-union ratio. This difference is calculated as the loss. Loss function, when the two boxes overlap, If the value is 1, The value is 0; when the two boxes have no intersection, The value of is -1, The value is 2, so , the predicted box and the true box The closer it is to 0, the closer the morphological distance between the two frames is, otherwise the distance and similarity between the two frames are lower.

[0027] In this embodiment, the specific design steps of the loss function are as follows: Step 1: The tracker is trained in an end-to-end manner, combining L1 loss and KIoU loss, and the loss function is: (5) in, and Represent the basic value and prediction box respectively, , is a hyperparameter; Step 2: Joint learning of localization and classification may lead to suboptimal solutions for both tasks. This helps to decouple localization and classification. Therefore, the training process is divided into two stages, with localization as the main task and classification as the secondary task. In the first stage, except for the scoring head, the entire network only trains the localization-related loss in equation (5). In this stage, all search images contain the target object and the model learns the localization ability. In the second stage, only the scoring head is optimized, and the binary cross entropy loss is defined as: (6) in, For the base tag, is the prediction confidence.

[0028] In this embodiment, the specific steps of updating the target template are: Step 1: During the tracking process, in each frame, the target template is cropped from the current image and then input into the backbone for feature extraction. The network takes a search area from the current frame as input, determines whether to update the target template through the target perception module, and returns the predicted box as the final result without using any post-processing such as cosine window or bounding box smoothing. Step 2: By effectively utilizing the target's spatiotemporal context interaction information, the target template is dynamically updated to avoid tracking drift caused by complex environmental factors in UAV application scenarios and further improve the robustness of tracking.

[0029] In this embodiment, Figure 3 and Figure 4 As shown, a "flat" structure is adopted (see Figure 3 The module does not change the channel dimension. At this time, the attention part occupies Figure 4 The main computations in the “flat” Transformer model in , achieve comparable performance with slightly more computation, which can be easily reduced through further optimization.

[0030] In this embodiment, Figure 5 and Figure 6 As shown in the figure, the weight distribution of traditional attention and LSDA attention is given respectively. The weight w describes the relationship between the source sentence and the target sentence. The darker the color, the larger the weight wij, indicating that the i-th word in the source sentence pays more attention to the j-th word in the target sentence, and the attention map usually has a strong pattern: sparse and diagonal. They represent the relationship between certain specific words: sparse for long-term information and diagonal for correlation within a small neighborhood. The former is called a "global" relationship and the latter is called a "local" relationship.

[0031] In this embodiment, during the target tracking process of the model, the target template changes dynamically. In order to ensure the robustness of the tracking, it is very necessary to dynamically capture the temporal and spatial change information of the target and update the template in a timely manner. For this reason, a target state perception module is designed to realize the update of the target template. For the update of the template, during the tracking process, sometimes the template cannot be updated. For example, when the target is blocked or disappears, the target information of the current frame cannot be used for updating. To this end, it is first determined that the template is updated only when the search area contains the target; secondly, a prediction head is added to predict whether to perform an update operation. Specifically: during the reasoning process, two templates and corresponding features are initialized in the first frame. Then a search area is cropped and input into the network to generate a bounding box and a target confidence score. Only when the update interval is reached and the confidence score is higher than the threshold In order to accurately perceive the dynamic changes of the target model, the target state perception module will dynamically adjust the threshold according to the average confidence score of the previous frames. , the specific situation is: when the target is less disturbed by the outside world and the target state is continuously stable, the confidence score is high. The setting value of is also relatively high; when the target is greatly disturbed by the outside world, such as when the target is blocked or deformed, the target state is unstable and the confidence output value is small. At this time, the threshold is also lowered accordingly. To improve efficiency, the target state perception module sets the update interval to frames, and along with the threshold Dynamically adjust according to changes in

[0032] Although the specific embodiments of the present invention are described in detail in conjunction with the accompanying drawings, it should not be understood as limiting the scope of protection of the present invention. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still belong to the scope of protection of the present invention.

Claims

1. A method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention, comprising building a tracking model, tracking a target by the model, and updating a target template, characterized in that: The construction of the tracking model includes constructing a backbone network, performing a bottleneck structure "flattening" processing on the feature map output from the backbone network before inputting it into the backbone network, inputting the feature sequence into the LSDA module, obtaining the global and local information of the feature sequence, the decoder obtaining a target query and an enhanced feature sequence from the encoder as input, and sending the input to the frame prediction head to predict the object frame coordinates, the model tracking the target includes framing the location of the target object, and then fine-tuning the regression frame, thereby training the model to predict the target position and adding the intersection and union ratio of key point distances as a frame regression function to the TLSDA algorithm, and the target template update includes designing a target state perception module and updating the target template.

2. A method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention as claimed in claim 1, characterized in that: The specific steps of constructing the backbone network are as follows: Step 1: The input of the backbone network is a triplet of images: the template image of the initial target object , Search area of ​​the current frame and dynamic templates ; Step 2: After passing the main trunk, the template z , Search images x and dynamic templates t are mapped to , and , where the initial template provides spatial information, and the dynamic template can capture the changes in target appearance over time, thereby providing additional temporal information.

3. The method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention according to claim 1, characterized in that: The bottleneck structure is "flattened" into a "flat" structure. Without changing the channel dimension, attention occupies the main calculation in the "flat" Transformer model, achieving comparable performance under large computational load, which can be reduced through further optimization.

4. The method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention according to claim 1, characterized in that: The specific steps of the LSDA module are as follows: Step 1: Capture global and local contexts separately. The LSDA module follows a two-branch design, where the left branch captures the global context, while the right branch models the local context. Step 2: Split the features into two parts along the channel dimension and mix them by the following FFN layer, which reduces the overall computational effort by half. Step 3: The left branch is a normal attention module, but the channel dimension is reduced by half. For the right branch that captures local relations, a natural idea is to apply convolution on the sequence. Step 4: By sliding the window, the diagonal line is covered; Step 5: To further reduce the amount of computation, replace the ordinary convolution with a lighter version consisting of a linear layer and a depth-wise convolution. Step 6: By placing the attention and convolution modules side by side, the model is encouraged to process the feature sequence from two different perspectives: global and local, so that the architecture can benefit from focus.

5. The method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention according to claim 1, characterized in that: The decoder is different from DETR which uses 100 object queries. The model only inputs one query into the decoder to predict a bounding box of the target object. There is only one prediction. The Hungarian algorithm used to predict associations in DETR is deleted. The decoder stacks N decoder layers, each of which consists of a self-attention parallel convolution, an encoder-decoder attention, and a feedforward network. In the Encoder-Decoder attention module, the target query can focus on all positions on the template and search area features, thereby learning a robust representation of the final bounding box prediction.

6. The method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention according to claim 1, characterized in that: The specific steps of the border prediction head are as follows: Step 1: Extract search area features from the output sequence of the encoder; Step 2: Calculate the similarity between the search region features and the decoder’s output embedding; Step 3: Multiply the similarity score by the feature element of the search area to enhance the important areas and weaken the areas with poor discriminability, and reshape the new feature sequence into The feature map of Step 4: Input a simple fully convolutional network FCN, which consists of L stacked Conv-BN-ReLU layers, and outputs two probability maps for the upper left corner and lower right corner of the object bounding box respectively. and ,By estimating the probability distribution of border corner points, the ,quality of border estimation is improved.

7. The method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention according to claim 1, characterized in that: The specific calculation steps of the key point distance intersection are as follows: Step 1: Frame and frame They are the real box and the predicted box respectively, and the line segments between the corresponding vertices of the two boxes are recorded , , The length is , , , remember the line segment between the center points of the two frames The length is , respectively extend the length of the two frames to obtain the minimum covering frame, and connect the points and Point , get the diagonal of the minimum covering box , the length is , The calculation is done by Adding a penalty term on the basis of: （1） in, , to ensure Non-negativity of values, penalty term It must be a value between [0, 1]; Step 2: Calculate the sum of the squares of the Euclidean distances between four groups of points, including three pairs of vertices and one pair of center points. Then, normalize the sum to ensure the normalization of the penalty term. The normalized benchmark meets two conditions: (1) The normalized benchmark is used as the denominator to compare with the sum, so the benchmark cannot be 0; (2) To ensure that the penalty term is a value distributed between [0, 1], the normalized benchmark must be larger than the value of the above sum, and the diagonal length of the minimum covering box between the predicted box and the true box is selected. c is the normalized benchmark, c The minimum value of is the diagonal length of the real box, which satisfies condition 1; in addition, the diagonal length of the minimum coverage box c It is the longest distance between any two points in the two boxes, which is greater than the distance between the three vertices and the center point corresponding to the two boxes, satisfying condition 2. In addition, the value obtained by using the Euclidean square of the distance between points is larger than that obtained by directly using the distance between points, which improves the calculation sensitivity of the penalty term. The calculation formula of the penalty term is: （2） in, , , , , Respectively represent , , , and The square of the Euclidean distance between the two boxes, the numerator is the Euclidean distance between the four groups of corresponding key points of the two boxes, obtained by formula (1) and formula (2): The calculation formula is: （3） Step 3: From formula (3), we can get: (1) ,when The value is 1, and the penalty value is 0, that is, when the two frames overlap, The value is 1; when The value is 0, and the penalty value is 1, that is, when the two boxes have no intersection, The value is -1, (2) due to It is a range value, not sensitive to scale, and will not change due to image scale issues. Value (3) When the two boxes do not intersect, Using the distance between points as gradient optimization makes up for The shortcomings, The loss function is: （4） Step 4: The loss function uses the ideal intersection-union ratio minus the actual calculated intersection-union ratio. This difference is calculated as the loss. Loss function, when the two boxes overlap, If the value is 1, The value is 0; when the two boxes have no intersection, The value of is -1, The value is 2, so , the predicted box and the true box The closer it is to 0, the closer the morphological distance between the two frames is, otherwise the distance and similarity between the two frames are lower.

8. The method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention according to claim 7, characterized in that: The specific design steps of the loss function are as follows: Step 1: The tracker is trained in an end-to-end manner, combining L1 loss and Loss, the loss function is: （5） in, and Represent the basic value and prediction box respectively, , is a hyperparameter; Step 2: Joint learning of localization and classification may lead to suboptimal solutions for both tasks. This helps to decouple localization and classification. Therefore, the training process is divided into two stages, with localization as the main task and classification as the secondary task. In the first stage, except for the scoring head, the entire network only trains the localization-related loss in equation (5). In this stage, all search images contain the target object and the model learns the localization ability. In the second stage, only the scoring head is optimized, and the binary cross entropy loss is defined as: （6） in, For the base tag, is the prediction confidence.

9. The method for tracking a drone target based on spatiotemporal interaction and long- and short-term attention according to claim 1, characterized in that: The specific steps of updating the target template are as follows: Step 1: During the tracking process, in each frame, the target template is cropped from the current image and then input into the backbone for feature extraction. The network takes a search area from the current frame as input, determines whether to update the target template through the target perception module, and returns the predicted box as the final result without using any post-processing such as cosine window or bounding box smoothing. Step 2: By effectively utilizing the target's spatiotemporal context interaction information, the target template is dynamically updated to avoid tracking drift caused by complex environmental factors in UAV application scenarios and further improve the robustness of tracking.

Images