Aviation ground target tracking method based on cross-space-time response aggregation fusion

Abstract

The invention relates to a cross-space-time response aggregation fusion aviation ground target tracking method, and relates to the technical field of computer vision. The problems of low resolution, small target, similar target interference and the like in an aviation ground scene in the prior art are solved, and the technical problem that an existing twin neural network ignores time context information of continuous frames and is difficult to realize accurate and robust tracking is solved. The aviation ground target tracking method comprises a feature extraction network, a feature fusion network, an enhanced channel attention network and an anchor-frame-free classification regression prediction network. According to the aviation ground target tracking method based on cross-space-time response aggregation fusion, the speed is considered at the same time, the provided self-adaptive time context aggregation module ingeniously utilizes the calculation results of the first several frames, and the tracking performance is improved while the additional calculation amount is hardly increased.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and particularly to an aerial ground target tracking method for cross-time and space response aggregation and fusion. Background Art

[0002] Object tracking is a core task in the field of computer vision, aiming to analyze the object information in a video sequence and track the dynamic changes of the object in real time. The tracker usually obtains the initial position and appearance features of the object from the first frame of the video sequence and predicts the movement trajectory of the object in subsequent frames by calculation. Object tracking technology is widely used in various scenarios, including autonomous driving, video surveillance, behavior analysis, etc. With the diversification of application scenarios, object tracking technology faces increasingly complex challenges. Especially in a dynamic environment, how to improve the accuracy and robustness of tracking remains a research hotspot. Aerial ground tracking, as an important branch of object tracking, has important application value in the fields of remote sensing detection, environmental monitoring, military reconnaissance, etc. Especially in the fields of aerial photography, intelligent monitoring, border patrol, post-disaster assessment, traffic management, etc., the aerial ground tracking technology can provide high-resolution real-time image and video data to help decision-makers quickly obtain the dynamic information of ground targets. However, although the aerial ground tracking technology has made certain progress in these fields, there are still many technical problems to be solved. Especially compared with traditional object tracking, the challenges faced by aerial ground tracking are more complex and diverse.

[0003] First, a significant challenge in aerial ground tracking is the low-resolution and small-target problem. Aerial platforms (such as drones, satellites) are usually at relatively high flight altitudes, resulting in relatively low-resolution images taken. The low-resolution images make the details and texture information of the target unclear. Especially when shooting at a long distance or at a high altitude, the target has fewer pixels and low distinctiveness, which makes it difficult for traditional feature extraction methods to accurately capture the target features. In addition, the size of the target in aerial ground tracking is usually small, and the appearance of the target is often easily interfered by the surrounding background or similar objects, further increasing the difficulty of tracking. Second, occlusion, scale variation, and size change are also important challenges in aerial ground tracking. Due to the constantly changing motion trajectory and perspective of the aerial platform, the features such as the appearance, size, and shape of the target will change significantly over time. When the target is occluded or enters the edge of the field of view, traditional tracking algorithms often have difficulty maintaining stable tracking of the target. In addition, factors such as the scale change and pose change of the target will also cause errors in the feature extraction process, thus affecting the tracking accuracy. In addition, since aerial ground tracking usually requires real-time calculations on platforms with limited resources, how to balance the consumption of computing resources while ensuring tracking accuracy remains an important research topic. The computing power and storage capacity of aerial platforms are often limited, especially on devices such as drones. Therefore, how to design an efficient tracking algorithm to achieve efficient target tracking with limited computing resources has become a difficult point in this field.

[0004] To address the above challenges, many researchers have introduced deep learning methods into aerial ground target tracking. Deep learning, especially convolutional neural networks (CNNs) and siamese neural networks, has demonstrated powerful capabilities in image processing and target recognition, and can effectively extract high-level features in images. The siamese neural network, through its structure design of sharing weights, can efficiently perform target feature matching when dealing with target tracking tasks. Especially in the case of target appearance changes and occlusions, it can maintain strong robustness. Compared with traditional tracking methods based on handcrafted features, the advantages of the siamese neural network in aerial ground tracking are particularly prominent. It can automatically learn features suitable for the current task through a deep network, thus avoiding the complexity and limitations of manual feature selection. Specifically, when dealing with aerial ground target tracking, the siamese neural network can effectively handle the scale change, pose change, and background interference of the target, thereby improving the accuracy and stability of target tracking.

[0005] In addition to feature extraction, how to utilize temporal context information is also crucial for addressing the challenges in aerial-to-ground tracking. The temporal context information of consecutive frames is very important for accurate tracking. Especially in situations such as low resolution, small targets, and similar interferences, the information of a single frame is often insufficient to extract sufficient discriminative features. In such cases, the motion information between consecutive frames can provide important clues for target localization and motion trajectories, thereby significantly improving the robustness and accuracy of tracking. However, many existing aerial-to-ground tracking methods often neglect the role of temporal context information and mainly rely on single-frame images and target information of the current frame. Although this method works well in simple scenarios, its tracking performance is often limited in complex scenarios (such as occlusion, rapid target changes, etc.). Therefore, future research on aerial-to-ground tracking should pay more attention to utilizing the context information in the time series, combining the information of historical frames and the current frame, establishing a more accurate target motion model, and then enhancing the stability and accuracy of tracking. Finally, since aerial-to-ground tracking needs to run on platforms with limited computing resources, how to reduce the computational overhead and improve the real-time performance of the algorithm while ensuring the tracking accuracy is also an urgent problem to be solved currently. Traditional deep learning models usually have large computational requirements and are difficult to operate efficiently on resource-limited unmanned aerial vehicle or satellite platforms. To address this issue, researchers have proposed some optimization strategies, such as model compression, quantization, pruning, etc., to reduce the size and computational complexity of the model, thereby achieving more efficient tracking performance.

[0006] Siamese Network is widely used in target tracking tasks. It processes the target frame and the search frame respectively through two feature extraction networks with shared weights to extract their feature maps, and uses deep cross-correlation operations for similarity matching to finally find the position of the target in the search area. This method has achieved good results in some simple tracking scenarios, especially under the conditions of stable target appearance and simple background. However, facing complex aerial-to-ground target tracking tasks, existing Siamese Network-based technologies have exposed many drawbacks, especially in environments with low resolution, small targets, and interference from similar objects.

[0007] First of all, existing Siamese Network methods ignore the temporal context information between consecutive frames. In actual target tracking, video sequences usually contain rich dynamic information between consecutive frames, which is crucial for the precise localization and tracking of targets. However, traditional Siamese networks only focus on the similarity matching between the target frame and the search frame and completely ignore the temporal relationship between frames. At this time, although Siamese Network can effectively extract features in a single frame, it does not fully utilize the motion trajectories and appearance changes of the target in multiple consecutive frames, making the model often lack sufficient robustness when dealing with complex dynamic scenarios.

[0008] Secondly, challenges such as low resolution, small targets, and complex backgrounds are often faced in aerial-to-ground target tracking. In the images captured by aerial platforms, the number of pixels of the target is relatively small, and the distinguishability from the background or similar objects is low. Existing deep learning methods, even deep neural networks, are difficult to extract strongly discriminative features. At this time, traditional Siamese neural network methods are difficult to effectively distinguish targets from the background in low-quality images, resulting in a decrease in tracking accuracy. In addition, aerial-to-ground tracking often faces problems such as rapid movement, occlusion, and scale changes of targets. These dynamic changes increase the difficulty of target tracking, and the Siamese network lacks an effective mechanism to handle these changes. More seriously, existing tracking methods do not effectively handle the occlusion problem of targets. In the aerial-to-ground scenario, the target is often partially occluded or disappears for a short time, and traditional Siamese network methods do not introduce information exchange of consecutive frames. Therefore, when the target is occluded, the model cannot recover the target position from the front and back frames, resulting in tracking failure. Summary of the Invention

[0009] Aiming at the problems of low resolution, small targets, interference from similar targets, etc. in the aerial-to-ground scenario in the prior art, and the technical problem that the existing Siamese neural network ignores the temporal context information of consecutive frames and is difficult to achieve accurate and robust tracking, the present invention provides an aerial-to-ground target tracking method based on cross-temporal response aggregation and fusion.

[0010] In order to solve the above technical problems, the technical solution of the present invention is specifically as follows:

[0011] An aerial-to-ground target tracking method based on cross-temporal response aggregation and fusion, comprising the following steps:

[0012] Step 1: Divide the initial frame in the collected aerial-to-ground tracking sequence into a target template frame, search for the sequence of the template frame with the current frame, and obtain the target template and the search area by cropping and normalizing the image frame.

[0013] Step 2: Input the target template and the search area after cropping and normalizing the image frame into the feature extraction network and send them into the cross-temporal interactive vision self-attention module to obtain the pyramid feature map. The feature map of the initial frame is the target template feature map, and the feature map of the current frame is the search area feature map.

[0014] Step 3: Perform deep feature fusion on the target template feature map and the search area feature map obtained in Step 2 to obtain a fused feature map, and then perform deep fusion on the obtained result and the search area feature map obtained in Step 2 again to obtain a fused feature response map.

[0015] Step 4: Input the response map output in Step 3 into the response saliency filtering aggregation network to perform feature saliency aggregation operation to obtain a predicted response map.

[0016] Step 5: Input the predicted response map into the anchor-free classification and regression network to obtain the final tracking result.

[0017] In the above technical solution, in step 2 of the aerial-to-ground target tracking method, AlexNet is selected as the network structure for feature extraction.

[0018] In the above technical solution, in step 3 of the aerial-to-ground target tracking method,

[0019] The cascaded self-attention encoder is:

[0020] ICIT(Q, K, V) = Norm(I + MLP(I))

[0021] I = Norm(Q + MultiHead(Q, Norm&AvgPool(K, V)))

[0022] P i = P i (Q4, K i , V i ), i ∈ 3, 4, 5

[0023] R i = ICIT(P i , P i , P i ), i ∈ 3, 4, 5

[0024]

[0025] R' = ICIT(R, R, R)

[0026] R'' = ICIT(R', R', R')

[0027] Among them, Q and Q i represent query variables, K and K i , V and V i correspond to the key and value in the input key-value pair respectively, which corresponds to the triple input in the ICIT module structure, Norm is the normalization layer operation, AvgPool is the average pooling, MLP is the fully connected layer, I is the intermediate variable of the residual structure, P i is the feature of the pyramid structure of low-level information and high-level semantics, and R is the result after fusion by the ICIT module.

[0028] In the above technical solution, step 4 of the aerial-to-ground target tracking method is specifically:

[0029] First, define the fused feature response map obtained in Step 3 as X. Perform global average pooling on the spatial dimension to compress each channel into a single value, resulting in a vector of size C×1×1. Pass the obtained vector through a simple one-dimensional convolution to generate channel attention weights. The size of the weight vector is also C×1×1, and the size of its convolution kernel k is adaptively determined:

[0030]

[0031] Among them, γ and b are hyperparameters. Normalize this weight vector through a Sigmoid function to obtain the final response saliency weight W:

[0032] W = Sigmoid(conv k (GAP(X)))

[0033] Among them, GAP represents the global average pooling operation, X is the fused feature response map, and conv k is the convolution operation with k as the convolution kernel.

[0034] Then, multiply the original fused feature response map and the obtained response saliency weight channel by channel;

[0035] After that, directly add and fuse the obtained result with the original fused feature response map weighted to obtain the predicted response map O: O = W·X + ε·X; where ε is the weighted parameter.

[0036] In the above technical solution, Step 5 of the aviation ground target tracking method is specifically as follows: Adopt an anchor-free mechanism to perform pixel-level discrimination on the obtained predicted response map. The classification branch discriminates the foreground and background of the pixels, and the regression branch directly takes each pixel point of the response map as the center, and regresses and calculates the distances between the upper, lower, left, and right boundaries and the center, and draws the tracking box.

[0037] The present invention has the following beneficial effects:

[0038] The aviation ground target tracking method based on cross-time and space response aggregation fusion of the present invention proposes an adaptive temporal context module on the basis of the existing Siamese neural network. This module can be divided into a temporal context encoder and a temporal context decoder. First, the encoder combines the context information of the response map of the target frame and the response map of the adjacent frame to obtain temporal prior information, and the decoder realizes the temporal context enhancement of the response map of the target frame by splicing and fusing the prior information. Through the encoder and decoder structures, effectively introduce the inter-frame context information into the tracking process of the current frame adaptively, improve the tracking ability of the tracker to challenges such as low resolution, small targets, and interference from similar targets, and realize a robust and accurate aviation ground target tracking method.

[0039] The method for aerial ground target tracking with cross - space - time response aggregation and fusion of the present invention takes speed into account. The proposed adaptive temporal context aggregation module cleverly utilizes the results calculated in the previous several frames, improving the tracking performance while hardly increasing the extra computational amount. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments.

[0041] Figure 1 It is a schematic diagram of the overall architecture of the method for aerial ground target tracking with cross - space - time response aggregation and fusion of the present invention.

[0042] Figure 2 It is a schematic diagram of the cross - space - time interactive vision self - attention module.

[0043] Figure 3 It is a schematic diagram of the response saliency filtering aggregation network structure. SPECIFIC EMBODIMENTS

[0044] The inventive concept of the present invention is as follows:

[0045] The method for aerial ground target tracking with cross - space - time response aggregation and fusion of the present invention is a complete solution applied to target tracking in the aerial ground scenario.

[0046] The method for aerial ground target tracking with cross - space - time response aggregation and fusion of the present invention has four components: a feature extraction network, a similarity matching network, an adaptive temporal context module (adaptive temporal context aggregation module, ATCAM), and an anchor - free classification and regression prediction network.

[0047] The method for aerial ground target tracking with cross - space - time response aggregation and fusion of the present invention proposes an adaptive spatio - temporal context aggregation module. This module can be divided into a temporal context encoder and a temporal context decoder. First, the encoder combines the context information of the response map of the target frame and the response maps of adjacent frames to obtain temporal prior information, and the decoder realizes the temporal context enhancement of the response map of the target frame through splicing and fusing the prior information. During the encoding process, we choose the response map to extract spatio - temporal information. This is because the response map, as the result of the fusion of the template branch and the search region branch, contains the information of both branches of adjacent frames, rather than single template information or search region information. In addition, choosing the response map for temporal information encoding is also efficient compared with the feature extraction link, requiring less storage resources and computational resources.

[0048] The overall architecture of the method for aerial ground target tracking with cross - space - time response aggregation and fusion of the present invention is as Figure 1 shown, and includes the following steps:

[0049] Step 1: Divide the initial frame in the collected aerial-to-ground tracking sequence into a target template frame, search for the sequence of the current frame in the template frame, and perform cropping and normalization processing on the image frame to obtain a target template and a search area.

[0050] Step 2: Input the target template and the search area after cropping and normalization processing of the image frame into a feature extraction network and send them to a spatio-temporal interactive vision self-attention module to obtain a pyramid feature map. The feature map of the initial frame is the target template feature map, and the feature map of the current frame is the search area feature map.

[0051] Step 3: Perform deep feature fusion on the target template feature map and the search area feature map obtained in Step 2 to obtain a fused feature map, and then perform deep fusion on the obtained result and the search area feature map obtained in Step 2 again to obtain a fused feature response map.

[0052] Step 4: Input the fused feature response map output in Step 3 into a response saliency filtering aggregation network to perform feature saliency aggregation operation to obtain a prediction response map.

[0053] Step 5: Input the prediction response map into an anchor-free classification regression network to obtain the final tracking result.

[0054] The present invention will be described in detail below with reference to the accompanying drawings.

[0055] As Figure 1 shown, the aerial-to-ground target tracking method with spatio-temporal response aggregation fusion of the present invention consists of four parts: a feature extraction network, a feature fusion network, a strengthened channel attention network, and an anchor-free classification regression prediction network.

[0056] First, Step 1 is to process the aerial remote sensing image sequence. The initial frame is divided into a target template frame Z, and the current frame sequence is used as search frames X in turn. The images are uniformly subjected to cropping and normalization processing to obtain a target template and a search template. The size of the target template is 127×127×3, and the size of the search area is 255×255×3.

[0057] For the feature extraction network in Step 2, in order to balance accuracy and the number of parameters of the algorithm, the present invention selects AlexNet as the network structure for feature extraction. AlexNet is a relatively lightweight convolutional neural feature extraction network in deep learning and performs particularly well in the field of computer vision. The AlexNet network structure can learn more complex features compared to other traditional convolutional neural networks, can better prevent the occurrence of overfitting phenomena, and significantly improve accuracy and computational efficiency. The target template F x and the search area Fz 。

[0058] In terms of feature fusion, the encoding layer of the traditional attention mechanism consists of a multi-head attention module and a multi-layer perceptron module. The attention function operates on the query Q, key K, and value V using a scaled dot-product model, i.e.:

[0059]

[0060] where C is the dimension of the key K of the normalized attention, Q represents the query variable, and K and V correspond to the key and value in the input key-value pair respectively. Pos is the positional encoding added to the input of each attention layer, and Softmax is the Softmax activation function.

[0061] We adopted a new type of cascade self-attention encoder (Cascade Transformer Encoder), named Transformer Cascade Fusion (TME), which uses an Intertemporal Correlation Interactive Vision Transformer (ICIT), as Figure 2 shown. The calculation process of the ICIT module can be summarized as follows:

[0062] ICIT(Q, K, V) = Norm(I + MLP(I))

[0063] I = Norm(Q + MultiHead(Q, Norm & AvgPool(K, V)))

[0064] where Q, K, and V are the ternary inputs, Norm is the normalization layer, AvgPool is the average pooling, MLP is the fully connected layer, and I is the intermediate variable of the residual structure.

[0065] Specifically, we utilized the features P of the pyramid structure of low-level information and high-level semantics i , i ∈

[0066] {3, 4, 5}. The high-level semantic information helps identify the category and overall features of the target, while the low-level information can capture details such as the fine texture and edges of the target. The configurations from bottom to top are P5, P4, and P3 respectively. The bottom-up path is the feed-forward convolution from the backbone architecture and generates the feature hierarchy {P5, P4, P3}. First, the information with pyramid features is fed into a 1×1 convolutional layer for dimension reduction and flattening. P3, P4, and P5 are provided as the key K and value V of their own feature hierarchies, and P4 is used as the query Q for the P3, P4, and P5 feature hierarchies. Based on the above conditions, the respective features are input into the ICIT module. After that, we take the results P obtained from the three features P3, P4, and P5 respectivelyi , for \(i\in\{3,4,5\}\), they are added and fused to obtain \(R\). Finally, the obtained fused result \(R\) is input into the ICIT module again to get \(R'\). The query \(Q\), key \(K\), and value \(V\) used are provided by itself, and this step is repeated again to finally obtain the semantic information \(R''\). In this way, we have completed the design of this new cascaded self-attention encoder. The above process can be summarized as follows:

[0067] P i =P i (Q4, K i , V i ), \(i\in\{3,4,5\}\)

[0068] R i =ICIT(P i , P i , P i ), \(i\in\{3,4,5\}\)

[0069]

[0070] R′ = ICIT(R, R, R)

[0071] R″ = ICIT(R′, R′, R′)

[0072] Among them, Q i represents the query variable, K i and V i correspond to the key and value in the input key-value pair respectively, P i is the feature of the pyramid structure of low-level information and high-level semantics, and \(R\) is the result after fusion by the ICIT module.

[0073] Then, the present invention designs the similarity matching between the template feature response map and the target feature response map. First, a depth cross-correlation operation is performed on the two to strengthen the extraction of the target feature; then, in order to strengthen the more refined detection and tracking of the target by the entire algorithm, the present invention designs a feed-forward residual multi-dimensional fusion edge for the result after the above depth cross-correlation operation, that is, fusing it with the target feature response map before fusion. Such an operation can make the entire algorithm have a deeper correlation between the target and the template, which is beneficial to improving the deeper detection and tracking of the target by the entire algorithm.

[0074] Then, in step four, the present invention processes the obtained semantic information through a response saliency filtering aggregation network. The network structure diagram of the response saliency filtering aggregation network is as Figure 3 shown. The specific approach is as follows:

[0075] First, perform global average pooling on the fused feature response map X obtained in the previous step in the spatial dimension, compress each channel into a single value, and obtain a vector X′ of size C×1×1; pass the obtained vector through a simple one-dimensional convolution to generate channel attention weights X″, and the size of the weight vector is also C×1×1, and the size of its convolution kernel k is adaptively determined:

[0076]

[0077] Among them, γ and b are hyperparameters. Pass this weight vector through a Sigmoid function for normalization to obtain the final response saliency weight W:

[0078] W = Sigmoid(conv k (GAP(X)))

[0079] Then, multiply the original fused feature response map and the obtained response saliency weight channel by channel, and complete the channel attention weighting operation on the fused feature response map.

[0080] After that, directly add and fuse the obtained result with the original fused feature response map weighted. In this step, the weighting parameter is ε, which is a learnable parameter and is used as a scalar weight to weight the input feature map, and the value of ε will be adjusted according to the gradient descent algorithm of the model to learn the most suitable weight for the model. According to the above operations, it is more convenient to learn deeper features, and finally obtain the prediction response map O.

[0081] O = W·X + ε·X

[0082] Finally, in step five, the present invention uses an anchor-free mechanism to perform pixel-level discrimination on the obtained prediction response map. The classification branch discriminates foreground and background for pixels, while the regression branch directly takes each pixel point of the response map as the center, and regresses and calculates the distances from the upper, lower, left, and right boundaries to the center, and draws the tracking box. The anchor-free classification and regression network avoids complex hyperparameter adjustment, reduces the computational amount brought by multiple anchor boxes, makes the model more concise and lightweight, and is more suitable for aerial-to-ground template tracking.

[0083] The test results on aerial-to-ground datasets such as UAV123 and DTB70 show that the accuracy of the aerial-to-ground target tracking method with cross-temporal and cross-spatial response aggregation fusion of the present invention is higher than that of existing tracking methods.

[0084] The aviation ground target tracking method for cross-time-and-space response aggregation and fusion of the present invention proposes an adaptive temporal context module on the basis of the existing twin neural network. This module can be divided into a temporal context encoder and a temporal context decoder. First, the encoder combines the context information of the response map of the target frame and the response maps of adjacent frames to obtain temporal prior information, and the decoder realizes the temporal context enhancement of the response map of the target frame by splicing and fusing the prior information. Through the encoder and decoder structures, the effective inter-frame context information is adaptively introduced into the tracking process of the current frame, improving the ability of the tracker to cope with challenges such as low resolution, small targets, and interference from similar targets, and realizing a robust and accurate aviation ground target tracking method.

[0085] The aviation ground target tracking method for cross-time-and-space response aggregation and fusion of the present invention takes speed into account at the same time. The proposed adaptive temporal context aggregation module cleverly uses the results calculated in the previous few frames, improving the tracking performance while hardly increasing the additional computational amount.

[0086] Obviously, the above embodiments are only examples given for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made on the basis of the above description. It is not necessary and impossible to enumerate all the implementation manners here. And the obvious changes or modifications derived therefrom still fall within the protection scope of the present invention.

Claims

1. An air-to-ground target tracking method for cross-time-and-space response aggregation and fusion, characterized in that, It includes the following steps: Step 1: Divide the initial frame in the collected aerial-to-ground tracking sequence into a target template frame, search for the template frame in the current frame sequence, and perform cropping and normalization processing on the image frame to obtain a target template and a search area; Step 2: Input the target template and search area after cropping and normalization processing of the image frame into a feature extraction network and send them into a cross-temporal and spatial interactive visual self-attention module to obtain a pyramid feature map. The feature map of the initial frame is the target template feature map, and the feature map of the current frame is the search area feature map; Step 3: Perform deep feature fusion on the target template feature map and the search area feature map obtained in Step 2 to obtain a fused feature map, and then perform deep fusion on the obtained result and the search area feature map obtained in Step 2 again to obtain a fused feature response map; Step 4: Input the fused feature response map output in Step 3 into a response saliency filtering aggregation network to perform feature saliency aggregation operation to obtain a predicted response map; Step 5: Input the predicted response map into an anchor-free classification and regression network to obtain the final tracking result.

2. The method for tracking an aerial ground target by cross-time-and-space response aggregation and fusion according to claim 1, wherein In Step 2, AlexNet is selected as the network structure for feature extraction.

3. The method for tracking an air-to-ground target by cross-time-and-space response aggregation and fusion according to claim 1, wherein In Step 3, the cascaded self-attention encoder is: ICIT(Q, K, V) = Norm(I + MLP(I)) I = Norm(Q + MultiHead(Q, Norm&AvgPool(K, V))) P i = P i (Q4, K i , V i ), i ∈ 3, 4, 5 R i = ICIT(P i , P i , P i ), i ∈ 3, 4, 5 R′ = ICIT(R, R, R) R″ = ICIT(R′, R′, R′) Among them, Q and Q i represent query variables, K and K i , V and V i correspond to the keys and values in the input key-value pairs respectively, which correspond to the triple input in the ICIT module structure, Norm is the normalization layer operation, AvgPool is the average pooling, MLP is the fully connected layer, I is the intermediate variable of the residual structure, P i is the feature of the pyramid structure of low-level information and high-level semantics, and R is the result after fusion by the ICIT module.

4. The method for tracking an air-to-ground target by cross-time-and-space response aggregation and fusion according to claim 1, wherein Step 4 is specifically as follows: First, define the fused feature response map obtained in Step 3 as X, perform global average pooling in the spatial dimension, compress each channel into a numerical value to obtain a vector of size C×1×1; pass the obtained vector through a simple one-dimensional convolution to generate channel attention weights, and the size of the weight vector is also C×1×1, and the size of its convolution kernel k is adaptively determined: where γ and b are hyperparameters, pass this weight vector through a Sigmoid function for normalization to obtain the final response saliency weight W: W = Sigmoid(conv k (GAP(X))) Among them, GAP represents the global average pooling operation, X is the fused feature response map, and conv k is the convolution operation with a convolution kernel of k; Then, multiply the original fused feature response map by the obtained response saliency weight channel by channel; After that, directly add and fuse the obtained result with the original fused feature response map weighted to obtain a predicted response map O: O = W·X + ε·X; where ε is the weighted parameter.

5. The method for tracking an aerial target for cross-time-and-space response aggregation and fusion according to claim 1, characterized in that Step 5 is specifically as follows: Adopt an anchor-free mechanism to perform pixel-level discrimination on the obtained predicted response map. The classification branch discriminates foreground and background for pixels, and the regression branch directly takes each pixel point of the response map as the center, regresses and calculates the distances from the upper, lower, left, and right boundaries to the center, and draws a tracking box.

Images