Lightweight unmanned aerial vehicle target tracking method and system based on dual cross-correlation fusion

Abstract

The invention discloses a lightweight unmanned aerial vehicle target tracking method and system based on dual cross-correlation fusion. The method comprises the following steps: acquiring a video stream of a target area; adding an initial target frame to a first frame image of the video stream to calibrate the initial position of a target, taking the first frame image as a template image, and taking other frame images of the video stream as search images; preprocessing the template image and the search image; inputting the preprocessed template image and the search image into a pre-trained target tracking model to obtain a predicted position of the target in a certain frame of image of the video stream, thereby realizing target tracking; a twin network is adopted as an overall framework, a backbone network is a lightweight network, the real-time performance of an algorithm is improved while the network parameter quantity and the calculation quantity are reduced, a dual cross-correlation fusion module is adopted, the nonlinearity of fusion features is enhanced, target feature information is highlighted, an anchor-free position prediction head is used, and the tracking speed is improved. And the network parameter quantity is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of unmanned aerial vehicle vision technology, and in particular relates to a lightweight unmanned aerial vehicle target tracking method and system based on dual cross-correlation fusion. Background Art

[0002] Object tracking refers to the use of computer vision technology to track a target in a video sequence in order to achieve comprehensive analysis of the target's position, size, movement, etc. In the field of computer vision and machine learning, single target tracking is an important research field, mainly used in video surveillance, autonomous driving, drone control, robot control, etc.

[0003] The application prospects of drones in the field of target tracking are very broad, and their unique advantages make them have important application potential in multiple industries. For example, drones can conduct real-time monitoring in a wide area, especially for difficult-to-reach or dangerous areas. Equipped with high-resolution cameras and intelligent algorithms, drones can automatically identify and track suspicious targets, such as intruders or abnormal events; after natural disasters (such as earthquakes, floods, fires, etc.), drones can be quickly deployed to track the situation in the disaster area in real time, identify and track the location of trapped people, and thus provide accurate target information for rescue operations. In the agricultural field, drones can be used to track the growth of crops, monitor soil moisture, analyze pests and diseases, etc. Through precise target tracking and data analysis, farmers can manage farmland more effectively and improve agricultural production efficiency. Summary of the invention

[0004] Purpose of the invention: In order to solve the problem that drones cannot accurately track targets due to insufficient computing power of onboard processing equipment, the present invention proposes a lightweight drone target tracking method and system based on dual cross-correlation fusion, which greatly improves the accuracy, success rate and efficiency of drone tracking, and achieves a performance balance between tracking accuracy and real-time performance.

[0005] Technical solution: A lightweight UAV target tracking method based on dual cross-correlation fusion, comprising the following steps:

[0006] Step 1: Get the video stream of the target area;

[0007] Step 2: Add an initial target frame to the first frame image of the video stream to calibrate the initial position of the target, use the first frame image as a template image, and use other frame images of the video stream as search images;

[0008] Step 3: preprocessing the template image and the search image to obtain the preprocessed template image and the search image;

[0009] Step 4: Input the preprocessed template image and search image into the pre-trained target tracking model to obtain the predicted position of the target in a certain frame of the video stream;

[0010] Step 5: Based on the predicted position of the target in a certain frame of the video stream, the target is tracked;

[0011] The target tracking model includes: a backbone network, a dual cross-correlation fusion network and an anchor-free position prediction head;

[0012] The backbone network is used to perform feature extraction on the preprocessed template image to obtain a template image feature map, and to perform feature extraction on the preprocessed search image to obtain a search image feature map;

[0013] The dual cross-correlation fusion network is used to match and fuse the global pixel similarity of the template image feature map and the search image feature map to obtain the final classification map and regression map;

[0014] The anchor-free position prediction head is used to obtain the predicted position of the target in a certain frame image of the video stream according to the final classification map and regression map.

[0015] Furthermore, the backbone network includes a template image feature extraction branch and a search image feature extraction branch. The template image feature extraction branch is used to perform feature extraction on the preprocessed template image to obtain a template image feature map; the search image feature extraction branch is used to perform feature extraction on the preprocessed search image to obtain a search image feature map.

[0016] Furthermore, the dual cross-correlation fusion network includes: a depth cross-correlation module, a pixel-by-pixel cross-correlation module, a space and channel fusion module, and a reverse activation module;

[0017] The template image feature map and the search image feature map are input into the pixel-by-pixel cross-correlation module, and the feature map P is obtained after the pixel-by-pixel correlation operation. The search image feature map and the cropped template image feature map are input into the depth cross-correlation module, and the feature map d is obtained after the depth cross-correlation operation. The feature map d is input into the space and channel fusion module to obtain the feature map x. scf , the feature map x scf Input into the reverse activation module to obtain the feature map x iab ; The feature map P passes through the SE layer and the 1×1 convolution layer in sequence and is combined with the feature map x iab Combined, as the output of the dual cross-correlation fusion network.

[0018] Furthermore, the deep cross-correlation module is used to sum the element-by-element products in the local window on each channel to obtain the correlation value of the local window, which is expressed as:

[0019]

[0020] In the formula, and They represent the template image feature map and the search image feature map respectively; Depth(*) represents the depth cross-correlation operation;

[0021] The pixel-by-pixel cross-correlation module is used to perform element-by-element product summation of the corresponding channels of the template image feature map and the search image feature map at each pixel position to obtain the correlation value of the pixel position, which is expressed as:

[0022]

[0023] Where Pixel(*) represents the pixel-by-pixel cross-correlation operation.

[0024] Furthermore, the spatial and channel fusion module consists of a 3×3 Conv-BN branch E 33 and a 1×1 Conv-BN branch E 11 Composition; During inference, a re-parameterized space and channel fusion module is used, and the re-parameterized space and channel fusion module is expressed as: Where φ represents the reparameterization transformation;

[0025] The feature map x scf It is expressed as:

[0026] x scf =φ(E 33 (x)+E 11 (x))+x

[0027] Where x corresponds to the feature map d.

[0028] Furthermore, the reverse activation module includes the first 1×1Conv-BN branch Non-linear Gelu activation function layer and the second 1×1 Conv-BN branch The feature map x output by the reverse activation module iab , expressed as:

[0029]

[0030] Where Gelu represents the nonlinear Gelu activation function.

[0031] Furthermore, in the process of pre-training the target tracking model, the loss function L total It is expressed as:

[0032] L total =λ1Lcls +λ2L reg

[0033] L reg =1-IoU

[0034] L cls =-(ylog(p)+(1-y)log(1-p))

[0035] Where, L cls represents the classification loss function, L reg represents the regression loss function; λ1 and λ2 represent the ratio of the loss weight of the loss function, which is a training parameter; IoU represents the intersection-over-union ratio between the target box predicted by the model and the real box, y represents the real label, and p represents the predicted probability.

[0036] The present invention discloses a lightweight UAV target tracking system based on dual cross-correlation fusion, comprising:

[0037] An image acquisition unit, used for acquiring a video stream of a target area;

[0038] An onboard processing unit, used to obtain a predicted position of the target based on a video stream of the target area acquired by the image acquisition unit;

[0039] A control unit, used to control the UAV to achieve target tracking based on the predicted position of the target;

[0040] Wherein, in the airborne processing unit, the video stream of the target area acquired by the image acquisition unit is used to obtain the predicted position of the target, which specifically includes:

[0041] Add an initial target frame to the first frame image of the video stream to calibrate the initial position of the target, use the first frame image as a template image, and use other frame images of the video stream as search images;

[0042] Preprocessing the template image and the search image to obtain a preprocessed template image and a search image;

[0043] The preprocessed template image and search image are input into the pre-trained target tracking model to obtain the predicted position of the target in a certain frame of the video stream;

[0044] The target tracking model includes: a backbone network, a dual cross-correlation fusion network and an anchor-free position prediction head;

[0045] The backbone network is used to perform feature extraction on the preprocessed template image to obtain a template image feature map, and to perform feature extraction on the preprocessed search image to obtain a search image feature map;

[0046] The dual cross-correlation fusion network is used to match and fuse the global pixel similarity of the template image feature map and the search image feature map to obtain the final classification map and regression map;

[0047] The anchor-free position prediction head is used to obtain the predicted position of the target in a certain frame image of the video stream according to the final classification map and regression map.

[0048] Furthermore, the backbone network includes a template image feature extraction branch and a search image feature extraction branch. The template image feature extraction branch is used to perform feature extraction on the preprocessed template image to obtain a template image feature map; the search image feature extraction branch is used to perform feature extraction on the preprocessed search image to obtain a search image feature map.

[0049] Furthermore, the dual cross-correlation fusion network includes: a depth cross-correlation module, a pixel-by-pixel cross-correlation module, a space and channel fusion module, and a reverse activation module;

[0050] The template image feature map and the search image feature map are input into the pixel-by-pixel cross-correlation module, and the feature map P is obtained after the pixel-by-pixel correlation operation. The search image feature map and the cropped template image feature map are input into the depth cross-correlation module, and the feature map d is obtained after the depth cross-correlation operation. The feature map d is input into the space and channel fusion module to obtain the feature map x. scf , the feature map x scf Input into the reverse activation module to obtain the feature map x iab ; The feature map P passes through the SE layer and the 1×1 convolution layer in sequence and is combined with the feature map x iab Combined, as the output of the dual cross-correlation fusion network.

[0051] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0052] (1) The present invention adopts a twin network as the overall framework, including a feature extraction backbone network, a dual cross-correlation fusion module and an anchor-free position prediction head. The backbone network used in the method of the present invention is a lightweight network. Compared with the huge model of traditional target tracking, it reduces the number of network parameters and the amount of calculation while improving the real-time performance of the algorithm. At the same time, it greatly reduces the computing power overhead during the model operation process, thereby improving the tracking speed;

[0053] (2) In order to effectively improve the performance of the model, the dual cross-correlation fusion module used in the method of the present invention can enhance the nonlinearity of the fusion features, highlight the target feature information, better distinguish the target from the background, and improve the model discrimination ability;

[0054] (3) The method of the present invention uses an anchor-free position prediction head, which greatly improves the tracking speed and reduces the number of network parameters;

[0055] (4) The method of the present invention designs a target tracking model based on a twin network for devices with limited computing power, such as drones. It solves the problem that drones cannot accurately track targets due to insufficient computing power of the onboard processing equipment of the drone platform, achieves a performance balance between tracking accuracy and real-time performance, and greatly improves the practical application value of drone target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a schematic diagram of the overall framework structure of a lightweight UAV target tracking method based on dual cross-correlation fusion proposed by the present invention;

[0057] Figure 2 It is the structural block diagram of the space and channel fusion module SCF;

[0058] Figure 3 It is a structural block diagram of the reverse activation module IAB;

[0059] Figure 4 A flowchart of the training method for a lightweight UAV target tracking model;

[0060] Figure 5 This is a real tracking effect diagram. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is now further described in conjunction with the accompanying drawings and embodiments.

[0062] like Figure 2 As shown, this embodiment discloses a lightweight UAV target tracking method based on dual cross-correlation fusion, which mainly includes the following steps:

[0063] Step 1: Use the camera or image sensor device on the drone to capture a video of the target area and transmit the video in real time to the onboard processing device;

[0064] Step 2: Initialize the target calibration of the video. The initial calibration of the target is to add an initial target frame to the first frame of the video manually or automatically through the target detection algorithm to calibrate the initial position of the target. The first frame of the image calibrated with the initial target frame is used as the template image, and each frame after the first frame will be used as the search image;

[0065] Step 3: Perform preprocessing of cropping, padding, scaling, jittering, flipping and normalization on the template image and the search image to obtain the template region image z = 3 × 127 × 127 and the search region image {x2, x3…, x n}=3×255×255, where n is the nth frame image;

[0066] Step 4: Process the template area image and the search area image in a lightweight UAV target tracking model based on dual cross-correlation fusion to predict the position of the target in the nth frame image. The specific operations include:

[0067] like Figure 1 As shown, the lightweight UAV target tracking model based on dual cross-correlation fusion proposed in this embodiment includes a lightweight backbone network, a dual cross-correlation fusion network and an anchor-free position prediction head;

[0068] The lightweight backbone network includes a template image feature extraction branch and a search image feature extraction branch. The template image feature extraction branch extracts features from the template region image z to obtain the first frame template image feature map. The search image feature extraction branch searches for the search area image x2, x3…, x n Perform feature extraction to obtain the image feature map for each frame after the first frame.

[0069] Among them, Figure 1 As shown in Figure 1, the dual cross-correlation fusion network includes a deep cross-correlation module, a pixel-by-pixel cross-correlation module, a space and channel fusion module SCF, and an inverse activation module IAB. The template image feature map and the search image feature map are input into the pixel-by-pixel cross-correlation module, and the feature map P is obtained after the pixel-by-pixel correlation operation; the search image feature map and the cropped template image feature map are input into the deep cross-correlation module, and the feature map d is obtained after the deep cross-correlation operation; the feature map d is input into the space and channel fusion module to obtain the feature map x scf , the feature map x scf Input into the reverse activation module to obtain the feature map x iab ; The feature map P passes through the SE layer and the 1×1 convolution layer in sequence and is combined with the feature map x iab Combined, as the output of the dual cross-correlation fusion network.

[0070] Among them, the deep cross-correlation module sums the element-by-element products in the local window on each channel to obtain the correlation value of the local window, which is expressed as:

[0071]

[0072] In the formula, and They represent the template feature map and the search image feature map respectively; Depth(*) represents the depth cross-correlation operation.

[0073] The pixel-by-pixel cross-correlation module focuses on the feature similarity at a specific pixel position and is calculated at a finer level than the depth cross-correlation module. At each pixel position, the corresponding channels of the two feature maps are summed element by element to obtain the correlation value of the pixel position; it is expressed as:

[0074]

[0075] Where Pixel(*) represents the pixel-by-pixel cross-correlation operation;

[0076] The spatial and channel fusion module SCF performs nonlinear fusion of local spatial information and channel information in the deep cross-correlation operation. Figure 2 As shown in the figure, the spatial and channel fusion module SCF consists of a 3×3 Conv-BN branch E 33 and a 1×1 Conv-BN branch E 11 Branch E 33 With branch E 11 The same dimensional transformation is performed to fuse the channel information. Through the structural reparameterization technology, the structure of the spatial and channel fusion module SCF can be equivalent to the convolution kernel, reducing the number of parameters. In the formula, φ represents the reparameterization transformation. At the same time, jump connections are performed to maintain the original fusion features. The expression of the entire space and channel fusion module SCF is:

[0077] Take the feature map d as the input of the following formula, that is, x, and get the corresponding feature map x scf ;

[0078]

[0079] like Figure 3 As shown in Fig. 2, the inverse activation module IAB includes two different 1×1Conv-BN branches, namely and And a nonlinear Gelu activation function, also using skip connections. The expression of the entire inverse activation module IAB is:

[0080]

[0081] Among them, the anchor-free position prediction head includes a classification regression module, a target prediction box calculation module and a loss function;

[0082] The classification regression module includes: a classification module and a regression module; the classification module uses the output of the dual cross-correlation fusion network as input to generate a two-channel classification map, one channel of the classification map represents the foreground probability, and the other channel represents the background probability. The regression module uses the output of the dual cross-correlation fusion network as input to generate a four-channel regression map. Each point on the regression map can be mapped to an area of ​​the search image, and each point represents the coordinates, height, and width of the prediction box.

[0083] The target prediction box calculation module screens the foreground probability on the classification map, takes the area with the largest foreground probability, sets the height and width of the prediction box in the corresponding area on the regression map, and obtains the final target prediction box.

[0084] The loss function uses IoU loss as the regression loss, expressed as: L reg =1-IoU, using cross entropy loss as classification loss, expressed as: L cls =-(ylog(p)+(1-y)log(1-p)). The overall loss function is derived from the classification loss function L cls And the regression loss function L reg Composition, expressed as:

[0085] L total =λ1L cls +λ2L reg

[0086] The lightweight UAV target tracking model based on dual cross-correlation fusion adopted in this embodiment is based on the 64-bit operating system Ubuntu 22.04, built using Python3.7 and PyTorch1.5.0 deep learning framework, and the hardware device used is Nvidia GeForce RTX 2080GPU. GOT10k is used as the video image in the training data set for training, the template branch input image size is 127×127, and the search branch input image size is 255×255. Specifically, the lightweight UAV target tracking model based on dual cross-correlation fusion includes: a feature extraction backbone network, a dual cross-correlation fusion module, and an anchor-free position prediction head.

[0087] The feature extraction backbone network uses a lightweight backbone network, specifically an improved MobileNetV3, and uses the parameters pre-trained on the ImageNet dataset to initialize the backbone network. Stochastic gradient descent (SGD) is used, the batch size is set to 64, and 150 epochs are trained. The initial learning rate of the training is 0.001, and the warm-up learning rate reaches 0.005 after 5 epochs, and then 0.005 is used as the initial learning rate for training. The learning rate decays exponentially from 0.005 to 0.0005 in the last 145 epochs.

[0088] The target tracking network is trained using stochastic gradient descent, and the IoU loss and cross entropy loss are calculated to update the network parameters. After reaching a maximum of 150 iterations, the training is stopped. The network parameters are saved.

[0089] After the training is completed, the training parameters are evaluated using offline tracking. The model is evaluated using two indicators: tracking accuracy and tracking success rate. The first indicator is to obtain the tracking accuracy by calculating the average percentage of frames in which the distance between the center point of the predicted target and the center point of the true target is less than or equal to the set threshold. In this embodiment, a threshold of 20 pixels is used as the accuracy score to sort the trackers. The tracking success rate is calculated by the average percentage of frames in which the intersection and union ratio of the predicted target bounding box and the true target bounding box is greater than the set threshold. The drone onboard processing equipment is used for target tracking, and the frame rate (FPS) is used to measure the real-time performance of the target tracking algorithm.

[0090] The trained model is evaluated on the UAV123 drone dataset. The UAV123 drone dataset is a dataset for a specific scenario. It is a video tracking dataset captured from a low-altitude drone. It contains 123 sequences with an average length of 1110 frames. It involves various types of targets such as people, vehicles, and animals. It also covers various complex scenes such as cities, suburbs, and indoors. In Table 1, Ours is the performance parameter of the model proposed in this embodiment, and the others are mainstream target tracking models. Succ. represents the success rate of the model, Prec. represents the accuracy of the model, and Avg. FPS represents the model tracking speed.

[0091] Table 1 Performance comparison

[0092] Tracker Succ. Prec Avg.FPS Ours 0.633 0.835 185 SiamBAN 0.631 0.833 45 SiamRPN++ 0.611 0.804 35 ABCNet 0.607 0.793 130 TCTrack 0.605 0.800 140 HIFT 0.589 0.787 160 SiamRPN 0.581 0.772 150 SiamDW 0.536 0.776 67 SiamFC 0.496 0.721 86

[0093] From the results in Table 1, it can be seen that the target tracking effect of the method in this embodiment is better than that of the mainstream target tracking model. In terms of the tracking success rate and accuracy indicators, the target tracking model proposed in this embodiment has achieved the best effect, and also improved the tracking speed. This shows that the method in this embodiment has a good application prospect and provides a new method and idea for the practical application of drones in target tracking in the future.

[0094] Step 5: Based on the predicted position, control the drone to track the target.

Claims

1. A lightweight UAV target tracking method based on dual cross-correlation fusion, characterized by: The following steps are involved: Step 1: Get the video stream of the target area; Step 2: Add an initial target frame to the first frame image of the video stream to calibrate the initial position of the target, use the first frame image as a template image, and use other frame images of the video stream as search images; Step 3: preprocessing the template image and the search image to obtain the preprocessed template image and the search image; Step 4: Input the preprocessed template image and search image into the pre-trained target tracking model to obtain the predicted position of the target in a certain frame of the video stream; Step 5: Based on the predicted position of the target in a certain frame of the video stream, the target is tracked; The target tracking model includes: a backbone network, a dual cross-correlation fusion network and an anchor-free position prediction head; The backbone network is used to perform feature extraction on the preprocessed template image to obtain a template image feature map, and to perform feature extraction on the preprocessed search image to obtain a search image feature map; The dual cross-correlation fusion network is used to match and fuse the global pixel similarity of the template image feature map and the search image feature map to obtain the final classification map and regression map; The anchor-free position prediction head is used to obtain the predicted position of the target in a certain frame image of the video stream according to the final classification map and regression map.

2. The lightweight UAV target tracking method based on dual cross-correlation fusion according to claim 1 is characterized in that: The backbone network includes a template image feature extraction branch and a search image feature extraction branch. The template image feature extraction branch is used to extract features from the preprocessed template image to obtain a template image feature map; the search image feature extraction branch is used to extract features from the preprocessed search image to obtain a search image feature map.

3. The lightweight UAV target tracking method based on dual cross-correlation fusion according to claim 1 is characterized in that: The dual cross-correlation fusion network includes: a depth cross-correlation module, a pixel-by-pixel cross-correlation module, a space and channel fusion module and a reverse activation module; The template image feature map and the search image feature map are input into the pixel-by-pixel cross-correlation module, and the feature map P is obtained after the pixel-by-pixel correlation operation. The search image feature map and the cropped template image feature map are input into the depth cross-correlation module, and the feature map d is obtained after the depth cross-correlation operation. The feature map d is input into the space and channel fusion module to obtain the feature map x. scf , the feature map x scf Input into the reverse activation module to obtain the feature map x iab ; The feature map P passes through the SE layer and the 1×1 convolution layer in sequence and is combined with the feature map x iab Combined, as the output of the dual cross-correlation fusion network.

4. The lightweight UAV target tracking method based on dual cross-correlation fusion according to claim 3 is characterized in that: The deep cross-correlation module is used to sum the element-by-element products in the local window on each channel to obtain the correlation value of the local window, which is expressed as: In the formula, and They represent the template image feature map and the search image feature map respectively; Depth(*) represents the depth cross-correlation operation; The pixel-by-pixel cross-correlation module is used to perform element-by-element product summation of the corresponding channels of the template image feature map and the search image feature map at each pixel position to obtain the correlation value of the pixel position, which is expressed as: Where Pixel(*) represents the pixel-by-pixel cross-correlation operation.

5. The lightweight UAV target tracking method based on dual cross-correlation fusion according to claim 3 is characterized in that: The spatial and channel fusion module consists of a 3×3 Conv-BN branch E 33 and a 1×1 Conv-BN branch E 11 Composition; During inference, a re-parameterized space and channel fusion module is used, and the re-parameterized space and channel fusion module is expressed as: Where φ represents the reparameterization transformation; The feature map x scf It is expressed as: x scf =φ(E 33 (x)+E 11 (x))+x Where x corresponds to the feature map d.

6. The lightweight UAV target tracking method based on dual cross-correlation fusion according to claim 3 is characterized by: The reverse activation module includes the first 1×1 Conv-BN branch Non-linear Gelu activation function layer and the second 1×1 Conv-BN branch The feature map x output by the reverse activation module iab , expressed as: Where Gelu represents the nonlinear Gelu activation function.

7. The lightweight UAV target tracking method based on dual cross-correlation fusion according to claim 1 is characterized in that: In the process of pre-training the target tracking model, the loss function L is used. total It is expressed as: L total =λ1L cls +λ2L reg L reg =1-IoU L cls =-(ylog(p)+(1-y)log(1-p)) Where, L cls represents the classification loss function, L reg represents the regression loss function; λ1 and λ2 represent the ratio of the loss weight of the loss function, which is a training parameter; IoU represents the intersection-over-union ratio between the target box predicted by the model and the real box, y represents the real label, and p represents the predicted probability.

8. A lightweight UAV target tracking system based on dual cross-correlation fusion, characterized by: include: An image acquisition unit, used for acquiring a video stream of a target area; An onboard processing unit, used to obtain a predicted position of the target based on a video stream of the target area acquired by the image acquisition unit; A control unit, used to control the UAV to achieve target tracking based on the predicted position of the target; Wherein, in the airborne processing unit, the video stream of the target area acquired by the image acquisition unit is used to obtain the predicted position of the target, which specifically includes: Add an initial target frame to the first frame image of the video stream to calibrate the initial position of the target, use the first frame image as a template image, and use other frame images of the video stream as search images; Preprocessing the template image and the search image to obtain a preprocessed template image and a search image; The preprocessed template image and search image are input into the pre-trained target tracking model to obtain the predicted position of the target in a certain frame of the video stream; The target tracking model includes: a backbone network, a dual cross-correlation fusion network and an anchor-free position prediction head; The backbone network is used to perform feature extraction on the preprocessed template image to obtain a template image feature map, and to perform feature extraction on the preprocessed search image to obtain a search image feature map; The dual cross-correlation fusion network is used to match and fuse the global pixel similarity of the template image feature map and the search image feature map to obtain the final classification map and regression map; The anchor-free position prediction head is used to obtain the predicted position of the target in a certain frame image of the video stream according to the final classification map and regression map.

9. The lightweight UAV target tracking system based on dual cross-correlation fusion according to claim 8, characterized in that: The backbone network includes a template image feature extraction branch and a search image feature extraction branch. The template image feature extraction branch is used to extract features from the preprocessed template image to obtain a template image feature map; the search image feature extraction branch is used to extract features from the preprocessed search image to obtain a search image feature map.

10. The lightweight UAV target tracking system based on dual cross-correlation fusion according to claim 8, characterized in that: The dual cross-correlation fusion network includes: a depth cross-correlation module, a pixel-by-pixel cross-correlation module, a space and channel fusion module and a reverse activation module; The template image feature map and the search image feature map are input into the pixel-by-pixel cross-correlation module, and the feature map P is obtained after the pixel-by-pixel correlation operation. The search image feature map and the cropped template image feature map are input into the depth cross-correlation module, and the feature map d is obtained after the depth cross-correlation operation. The feature map d is input into the space and channel fusion module to obtain the feature map x. scf , the feature map x scf Input into the reverse activation module to obtain the feature map x iab ; The feature map P passes through the SE layer and the 1×1 convolution layer in sequence and is combined with the feature map x iab Combined, as the output of the dual cross-correlation fusion network.

Images