Unmanned aerial vehicle lightweight single target tracking method and system, terminal and medium

Abstract

The embodiment of the invention relates to the field of data processing, and provides an unmanned aerial vehicle lightweight single-target tracking method and system, a terminal and a medium, and the method comprises the steps: extracting a first template frame from a first video, the first template frame being a first video frame in the first video under the view angle of a target unmanned aerial vehicle, and extracting a first search frame from a to-be-processed video frame, obtaining a first search frame set, wherein the to-be-processed video frame is a video frame except a first template frame in a first video; performing feature fusion on the first template frame and a first search frame in a first search frame set to obtain a first feature information set; performing feature enhancement on first feature information in the first feature information set to obtain a second feature information set; adjusting the initial tracking model to obtain a target tracking model; tracking a to-be-tracked target by using the target tracking model to obtain a target tracking result; and the accuracy of tracking the to-be-tracked target by the unmanned aerial vehicle is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a lightweight single-target tracking method, system, terminal, and medium for unmanned aerial vehicles (UAVs). Background Technology

[0002] UAV-based single object tracking refers to the task of continuously monitoring and locating specific targets (such as vehicles, pedestrians, and animals) using drones equipped with cameras. This technology has wide applications in fields such as military reconnaissance, security monitoring, traffic management, and ecological monitoring.

[0003] In the process of ecological monitoring, when monitoring and protecting some endangered wild animals, it is usually necessary to use drones to track the endangered wild animals that need to be protected, so as to determine their territory and living habits, and ensure that the scope of the protected area can be accurately determined, thereby effectively protecting endangered wild animals.

[0004] In recent years, with the rapid development of computer vision and deep learning technologies, existing tracking methods have begun to adopt generative or discriminative frameworks, updating the target model through online learning or iterative optimization. However, these methods generally rely only on sparse similarity measurements between template frames and search frames, ignoring continuous contextual information between adjacent frames, which is not conducive to capturing the dynamic changing trends of the target. For example, when tracking and filming Siberian tigers, existing tracking methods can only generate corresponding location scatter plots based on the location information of the tracked Siberian tiger, and determine the corresponding movement trajectory of the tracked Siberian tiger based on the scatter plot. However, obtaining the movement trajectory in this way cannot analyze the living habits of the tracked Siberian tiger, resulting in a lack of effective data support for protected area planning. Furthermore, when the Siberian tiger enters an area with lush vegetation, the drone may be unable to lock onto the tiger's location, thus preventing the drone from continuously tracking the tiger, resulting in low accuracy when the drone is tracking the target. Summary of the Invention

[0005] This application provides a lightweight single-target tracking method for unmanned aerial vehicles (UAVs), which can improve the accuracy of UAVs when tracking targets.

[0006] A first aspect of this application provides a lightweight single-target tracking method for unmanned aerial vehicles (UAVs), the method comprising: Extract a first template frame from the first video, where the first template frame is the first video frame in the first video from the perspective of the target drone; and extract a first search frame from the video frames to be processed to obtain a first search frame set, where the video frames to be processed are the video frames in the first video other than the first template frame. The first template frame and the first search frame in the first search frame set are fused to obtain the first feature information set. The first feature information in the first feature information set is enhanced to obtain the second feature information set; The initial tracking model is adjusted based on the second feature information in the second feature information set to obtain the target tracking model; The target tracking model is used to track the target and the target tracking result is obtained.

[0007] In this example, a first template frame is extracted from a first video, where the first template frame is the first video frame in the first video from the perspective of the target UAV. A first search frame set is obtained by extracting a first search frame from the video frames to be processed, where the video frames to be processed are all video frames in the first video except for the first template frame. The first template frame and the first search frame in the first search frame set are fused to obtain a first feature information set. The first feature information in the first feature information set is then enhanced to obtain a second feature information set. The initial tracking model is adjusted based on the second feature information in the second feature information set to obtain a target tracking model. The target tracking model is then used to track the target to be tracked to obtain the target tracking result, thereby improving the accuracy of the UAV in tracking the target to be tracked.

[0008] In one possible implementation, a method for fusing features between the first template frame and the first search frame in the first search frame set to obtain first feature information includes: The first template frame is subjected to size normalization processing to obtain the second template frame; The first search frame in the first search frame set is subjected to size normalization to obtain the second search frame set; The second search frame in the second template frame and the second search frame set are combined into a dual-frame image to obtain the first dual-frame image set; The first two-frame images in the first two-frame image set are subjected to deep encoding fusion to obtain the first mixed feature information set; The fused features corresponding to the second search frame in the second search frame set in the first mixed feature information are segmented to obtain the first feature information set.

[0009] In one possible implementation, a method for performing deep coding fusion on the first two-frame images in the first two-frame image set to obtain a first mixed feature information set includes: The first two-frame images in the first two-frame image set are segmented using preset segmentation pixel blocks to obtain m sets of first image blocks; where m represents the number of first two-frame images in the first two-frame image set. Linear projection is performed on the first image block in each set of first image blocks to obtain m sets of first image block vectors; The first image block vectors in the set of m first image block vectors are concatenated to obtain a first concatenated vector set. The first concatenated vector in the first concatenated vector set is input into the self-attention coding model for deep coding fusion to obtain the first mixed feature information set.

[0010] In one possible implementation, a method for enhancing the first feature information in the first feature information set to obtain a second feature information set includes: Construct the temporal correlation interaction matrix corresponding to each first feature in the first feature information set to obtain the first interaction matrix set; The first interaction matrix in the first interaction matrix set is weighted and fused element by element to obtain the first feature tensor information set. The first feature tensor information in the first feature tensor information set is subjected to dimensionality increase processing to obtain the second feature tensor information set; The second feature tensor information in the second feature tensor information set is reconstructed into a spatial feature map structure to obtain the second feature information set.

[0011] In one possible implementation, a method for adjusting an initial tracking model based on second feature information in the second feature information set to obtain a target tracking model includes: Extract the center score map of each second feature information in the second feature information set to obtain the first center score map set; Extract the size prediction map of each second feature information in the second feature information set to obtain the first size prediction map set; Extract the center offset map of each second feature information in the second feature information set to obtain the first center offset map set; The tracking box of the reference tracking target is predicted based on the first center score map in the first center score map set, the first size prediction map in the first size prediction map set, and the first center offset map in the first center offset map set, to obtain a target tracking box information set; The model loss value of the initial tracking model is calculated based on the target tracking boxes in the target tracking box information set to obtain the first loss value set; Determine whether all the first loss values ​​in the first loss value set are greater than a preset loss value threshold, and obtain the determination result; Based on the judgment results, the model parameters of the initial tracking model are adjusted using the backpropagation algorithm until the model loss value of the adjusted tracking model is less than or equal to the preset loss value threshold, thus obtaining the target tracking model.

[0012] A second aspect of this application provides a lightweight single-target tracking system for unmanned aerial vehicles (UAVs), the system comprising: The image processing unit is used to extract a first template frame from a first video, the first template frame being the first video frame in the first video from the perspective of the target drone, and to extract a first search frame from the video frames to be processed to obtain a first search frame set, wherein the video frames to be processed are video frames in the first video other than the first template frame. The extraction unit is used to perform feature fusion on the first template frame and the first search frame in the first search frame set to obtain a first feature information set. An enhancement unit is used to enhance the first feature information in the first feature information set to obtain a second feature information set. The adjustment unit is used to adjust the initial tracking model according to the second feature information in the second feature information set to obtain the target tracking model; The tracking unit is used to track the target to be tracked using the target tracking model and obtain the target tracking result.

[0013] In one possible implementation, the extraction unit is specifically used for: The first template frame is subjected to size normalization processing to obtain the second template frame; The first search frame in the first search frame set is subjected to size normalization to obtain the second search frame set; The second search frame in the second template frame and the second search frame set are combined into a dual-frame image to obtain the first dual-frame image set; The first two-frame images in the first two-frame image set are subjected to deep encoding fusion to obtain the first mixed feature information set; The fused features corresponding to the second search frame in the second search frame set in the first mixed feature information are segmented to obtain the first feature information set.

[0014] In one possible implementation, the extraction unit is specifically used for: The first two-frame images in the first two-frame image set are segmented using preset segmentation pixel blocks to obtain m sets of first image blocks; where m represents the number of first two-frame images in the first two-frame image set. Linear projection is performed on the first image block in each set of first image blocks to obtain m sets of first image block vectors; The first image block vectors in the set of m first image block vectors are concatenated to obtain a first concatenated vector set. The first concatenated vector in the first concatenated vector set is input into the self-attention coding model for deep coding fusion to obtain the first mixed feature information set.

[0015] In one possible implementation, the enhancement unit is specifically used for: Construct the temporal correlation interaction matrix corresponding to each first feature in the first feature information set to obtain the first interaction matrix set; The first interaction matrix in the first interaction matrix set is weighted and fused element by element to obtain the first feature tensor information set. The first feature tensor information in the first feature tensor information set is subjected to dimensionality increase processing to obtain the second feature tensor information set; The second feature tensor information in the second feature tensor information set is reconstructed into a spatial feature map structure to obtain the second feature information set.

[0016] In one possible implementation, the adjustment unit is specifically used for: Extract the center score map of each second feature information in the second feature information set to obtain the first center score map set; Extract the size prediction map of each second feature information in the second feature information set to obtain the first size prediction map set; Extract the center offset map of each second feature information in the second feature information set to obtain the first center offset map set; The tracking box of the reference tracking target is predicted based on the first center score map in the first center score map set, the first size prediction map in the first size prediction map set, and the first center offset map in the first center offset map set, to obtain a target tracking box information set; The model loss value of the initial tracking model is calculated based on the target tracking boxes in the target tracking box information set to obtain the first loss value set; Determine whether all the first loss values ​​in the first loss value set are greater than a preset loss value threshold, and obtain the determination result; Based on the judgment results, the model parameters of the initial tracking model are adjusted using the backpropagation algorithm until the model loss value of the adjusted tracking model is less than or equal to the preset loss value threshold, thus obtaining the target tracking model.

[0017] A third aspect of this application provides a terminal including a processor, an input device, an output device, and a memory, wherein the processor, input device, output device, and memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the step instructions as described in the first aspect of this application.

[0018] A fourth aspect of this application provides a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in the first aspect of this application.

[0019] A fifth aspect of this application provides a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in the first aspect of this application. The computer program product may be a software installation package. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This application provides a schematic diagram of the network architecture for a lightweight single-target tracking method for unmanned aerial vehicles (UAVs). Figure 2 This application provides a flowchart illustrating a lightweight single-target tracking method for unmanned aerial vehicles (UAVs). Figure 3 This is a schematic diagram of the structure of a terminal provided in an embodiment of this application; Figure 4 This application provides a schematic diagram of the structure of a lightweight single-target tracking system for unmanned aerial vehicles (UAVs). Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0024] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0025] Please see Figure 1 , Figure 1 This application provides a schematic diagram of the network architecture for a lightweight single-target tracking method for unmanned aerial vehicles (UAVs). The lightweight single-target tracking system includes a UAV 1, a target 2 to be tracked, and a server 3. The UAV can perform tracking tasks, such as tracking endangered wild animals. To ensure accurate acquisition of the endangered wild animal's habits and territory during tracking, continuous tracking is required.

[0026] After the drone detects endangered wild animals that need to be tracked, it continuously tracks the animals. Once in tracking mode, the drone continuously collects location and dynamic change information of the endangered wild animals and uploads the collected data to the server. After receiving the data uploaded by the drone, the server uses the tracking model stored on the server to determine the location of the endangered wild animals in the image. When the endangered wild animals are in areas with lush vegetation, the server predicts the location of the wild animals and instructs the drone to lock on, avoiding the inability of the drone to identify the endangered wild animals and thus enabling continuous tracking of endangered wild animals, thereby improving the accuracy of the drone in tracking the target.

[0027] To better understand the lightweight single-target tracking method for UAVs provided in this application, a brief introduction to existing lightweight single-target tracking methods for UAVs is given below. Existing methods typically employ generative or discriminative frameworks, updating the target model through online learning or iterative optimization. However, these methods generally rely solely on sparse similarity measurements between template frames and search frames, ignoring continuous contextual information between adjacent frames, which is detrimental to capturing the dynamic changing trends of the target. For example, when tracking and filming Siberian tigers, existing tracking methods can only generate corresponding location scatter plots based on the acquired tiger's location information and determine the tiger's trajectory based on the scatter plots. However, obtaining the trajectory in this way cannot reveal the tiger's habits, resulting in insufficient data support for protected area planning. Furthermore, when the tiger enters densely vegetated areas, the UAV may fail to lock onto its location, preventing continuous tracking and leading to low accuracy in tracking the target.

[0028] To address the aforementioned technical problems, this application provides a lightweight single-target tracking method for unmanned aerial vehicles (UAVs). This method involves extracting features from a first template frame and a first search frame in a first search frame set, enhancing the extracted feature information to obtain a second feature information set. The initial tracking model is then adjusted based on the second feature information in the second feature information set to obtain a target tracking model. This target tracking model is then used to track the target to obtain the target tracking result, thereby improving the accuracy of the UAV in tracking the target.

[0029] Please see Figure 2 , Figure 2 This application provides a flowchart illustrating a lightweight single-target tracking method for unmanned aerial vehicles (UAVs). Figure 2 As shown, the method includes: 101. Extract a first template frame from the first video, where the first template frame is the first video frame in the first video from the perspective of the target drone, and extract a first search frame from the video frames to be processed to obtain a first search frame set, where the video frames to be processed are video frames in the first video other than the first template frame.

[0030] Specifically, this can be achieved by acquiring the video image of the first frame in the first video to obtain the first video image; acquiring the contour information corresponding to the rectangular image region where the reference tracking target is located in the first video image to obtain the true target bounding box; cropping an image region with a size four times the area of ​​the true target bounding box from the first video image using the center point of the true target bounding box as a reference to obtain the first template frame; and cropping an image region with a size twice the area of ​​the true target bounding box from the video to be processed using the center point of the true target bounding box as a reference to obtain the first search frame set. The first video can be understood as a video from the drone's perspective when the drone is tracking the reference tracking target; the reference tracking target is the target being tracked by the drone in the first video.

[0031] 102. Perform feature fusion on the first template frame and the first search frame in the first search frame set to obtain the first feature information set.

[0032] Specifically, the process can involve standardizing the size of the first template frame and the first search frame in the first search frame set to obtain a second template frame and a second search frame set. The second template frame and the second search frame in the second search frame set are then combined to form a dual-frame image, resulting in a first dual-frame image set. The first dual-frame images in the first dual-frame image set are then segmented to obtain a set of m first image blocks. Vector representations corresponding to the first image blocks in the m first image block sets are obtained, resulting in a set of m first image block vectors. An improved template matching algorithm is used to concatenate the first image block vectors in the m first image block vector sets, and the concatenation is then input into a self-attention coding model for deep encoding fusion, resulting in a first mixed feature information set. The number of first mixed feature information items in the first mixed feature information set is m. The first mixed feature information in the first mixed feature information set is used to segment the fusion features corresponding to the second search frames in the second search frame set, resulting in a first feature information set. The first mixed feature information includes the fusion features corresponding to the second search frames in the second search frame set and the fusion features corresponding to the second template frame.

[0033] 103. Enhance the first feature information in the first feature information set to obtain the second feature information set.

[0034] Specifically, this can be achieved by constructing a temporal correlation interaction matrix corresponding to the first feature information in the first feature information set to obtain a first interaction matrix set; by performing element-wise weighted fusion of the first interaction matrix in the first interaction matrix set based on the first mixed feature information in the first mixed feature information set to obtain a first feature tensor information set; by performing dimensionality-up processing on the first feature tensor information in the first feature tensor information set to adjust the shape of the first feature tensor information in the first feature tensor information set to obtain a second feature tensor information set; and by reconstructing the second feature tensor information in the second feature tensor information set into a spatial feature map structure to obtain a second feature information set.

[0035] 104. Adjust the initial tracking model according to the second feature information in the second feature information set to obtain the target tracking model.

[0036] Specifically, this can be achieved by inputting the second feature information from the second feature information set into a lightweight multi-branch convolutional prediction network, using the center position prediction branch, size regression branch, and offset branch of the lightweight multi-branch convolutional prediction network to obtain the center position coordinates and width and height values ​​of the reference tracking target in the second search frame in the second search frame set, thus obtaining a target tracking box information set; calculating the loss value of the initial tracking model based on the target tracking box information in the target tracking box information set, thus obtaining a first loss value set; and adjusting the initial tracking model using a backpropagation algorithm based on a preset loss value threshold and the first loss value in the first loss value set, thus obtaining the target tracking model.

[0037] 105. Use the target tracking model to track the target and obtain the target tracking result.

[0038] One approach is to input the image corresponding to the target to be tracked into a target tracking model. The target tracking model can continuously track the target based on the input image and record the motion trajectory of the target to be tracked, thereby obtaining the target tracking result.

[0039] In this example, a first template frame is obtained by extracting a template frame from the video from the perspective of the target drone, and a first search frame is obtained by extracting a search frame from the video to be processed from the video from the perspective of the target drone. Feature extraction is performed on the first template frame and the first search frame to obtain first feature information. Feature enhancement is performed on the first feature information to obtain second feature information. The model parameters of the initial tracking model are adjusted according to the second feature information to obtain a target tracking model. The target tracking model is used to track the target to be tracked to obtain the target tracking result, thereby improving the accuracy of tracking the target when using a drone.

[0040] In one possible implementation, a method for fusing features of the first template frame and the first search frame in the first search frame set to obtain a first feature information set includes: A1. Perform size standardization processing on the first template frame to obtain the second template frame; A2. Perform size standardization processing on the first search frame in the first search frame set to obtain the second search frame set; A3. Combine the second search frame in the second template frame and the second search frame set into a dual-frame image to obtain the first dual-frame image set; A4. Perform deep coding fusion on the first two-frame images in the first two-frame image set to obtain the first mixed feature information set; A5. Segment the search frame features from the first mixed feature information to obtain the first feature information.

[0041] Specifically, the size of the first template frame can be scaled to a standard size of 128*128*3 using a common video image scaling method, thereby completing the size standardization process of the first template frame and obtaining the second template frame.

[0042] While obtaining the second template frame, the size of the first search frame in the first search frame set can be scaled to a standard size of 256*256*3 by using a common video image scaling method, thereby completing the size standardization process of the first search frame in the first search frame set and obtaining the second search frame set.

[0043] After obtaining the second template frame and the second search frame set, the first dual-frame image set can be obtained by determining the second template frame and the second search frame in the second search frame set as sub-images in the same dual-frame image.

[0044] After obtaining the first set of two-frame images, a second set of two-frame images can be obtained by acquiring the preceding two-frame image corresponding to each first two-frame image in the first set of two-frame images. The second two-frame image can be understood as a combination of a search frame from the video image preceding the video frame to be processed, after size normalization, and a second template frame. The first two frames in the first set of two-frame images are segmented to obtain m sets of first image blocks. The vector representations corresponding to the first image blocks in the m sets of first image blocks are obtained, resulting in m sets of first image block vectors. A common temporal token acquisition method is used to obtain the temporal token corresponding to each second two-frame image in the second set of two-frame images, resulting in a first set of temporal tokens. Based on the first temporal tokens in the first set of temporal tokens, deep encoding and fusion are performed on the first image block vectors in each set of first image block vectors to obtain a first set of hybrid feature information.

[0045] After obtaining the first set of mixed feature information, the fused features corresponding to the second search frame in the second search frame set in the first set of mixed feature information can be segmented using a general feature segmentation method to obtain the first set of feature information.

[0046] In one possible implementation, a method for performing deep coding fusion on the first two-frame images in the first two-frame image set to obtain a first mixed feature information set includes: B1. Use preset segmentation pixel blocks to segment the first double-frame image in the first double-frame image set to obtain m sets of first image blocks; where m represents the number of first double-frame images in the first double-frame image set; B2. Perform linear projection on the first image block in each set of first image blocks to obtain m sets of first image block vectors; B3. Perform vector concatenation on the first image block vectors in each of the first image block vector sets to obtain the first concatenated vector set; B4. Input the first concatenation vector in the first concatenation vector set into the self-attention coding model for deep coding fusion to obtain the first mixed feature information set.

[0047] Specifically, this can be achieved by selecting any one of the first two-frame images from the first two-frame image set to obtain a target first two-frame image; dividing the target first two-frame image into k non-overlapping image blocks using a preset segmentation pixel block of size 16x16 to obtain a first image block set; and then segmenting all the first two-frame images in the first two-frame image set using the same segmentation method as the target first two-frame image to obtain m first image block sets. Here, k represents the number of first image blocks in the first image block set.

[0048] After obtaining m sets of first image blocks, a target set of first image blocks can be obtained by arbitrarily selecting one set of first image blocks from the m sets of first image blocks; a set of first image block vectors can be obtained by linearly projecting the first image blocks in the target set of first image blocks using a general linear projection method; and a set of first image block vectors can be obtained by linearly projecting all the sets of first image blocks from the m sets of first image blocks using the same linear projection method as the target set of first image blocks.

[0049] After obtaining m sets of first image block vectors, the following steps can be taken: arbitrarily select one of the m sets of first image block vectors to obtain a target first image block vector set; use an improved template matching algorithm to extract the first image block vector corresponding to the second template frame in the target first image block vector set to obtain a second image block vector set; extract the first image block vector corresponding to the target second search frame in the target first image block vector set to obtain a third image block vector set; wherein, the target second search frame can be understood as the second search frame in the second search frame set that corresponds to the target first image block vector set; arrange the second image block vectors in the second image block vector set according to their corresponding temporal information to obtain a second image block vector sequence; arrange the third image block vectors in the third image block vector set according to their corresponding temporal information to obtain a third image block vector sequence. The second image block vectors in the second image block vector sequence are concatenated using a general vector concatenation method to obtain a first template frame vector. The third image block vectors in the third image block vector sequence are also concatenated using the same method to obtain a first search frame vector. The first template frame vector is then pooled using a general pooling method to obtain a second template frame vector. The first temporal token is added to the second template frame vector as a learnable classification token to obtain a third template frame vector. The method for adding the first temporal token to the second template frame vector set can be by numerically adding the first temporal token and the second template frame vector. The first template frame vector, the first search frame vector, and the third template frame vector are collectively determined as the first concatenated vector. The first image block vectors in all the first image block vector sets in the m first image block vector sets are concatenated using the same vector concatenation method as the target first image block vector set to obtain the first concatenated vector set.

[0050] After obtaining the first set of concatenated vectors, a target first concatenated vector can be obtained by arbitrarily selecting one first concatenated vector from the first set of concatenated vectors. The target first concatenated vector is then input into a self-attention encoding model, where a general layer normalization processing method is used to perform layer normalization processing on the first template frame vector, the first search frame vector, and the third template frame vector in the first concatenated vector to obtain first normalized feature information. The first normalized feature information is then subjected to linear projection processing to generate a query vector, a key vector, and a value vector corresponding to the first normalized feature information, resulting in a first query vector, a first key vector, and a first value vector. The first query vector and the first key vector are then normalized to obtain a second query vector and a second key vector. The attention score matrix of the first normalized feature information is calculated based on the second query vector, the second key vector, and the first value vector to obtain a first attention score matrix. A preset nonlinear correction term is added to the first attention score matrix to obtain a second attention score matrix. The second attention score matrix is ​​dimensionally restructured and linearly projected to extract attention output features, yielding first attention feature information. The first attention feature information is then subjected to layer scaling and random path discarding to obtain second attention feature information. A general residual connection method is used to perform a residual connection between the first normalized feature information and the second attention feature information to obtain second normalized feature information. This process is repeated n times: inputting the target first concatenation vector into the self-attention encoding model; performing layer normalization on the first template frame vector, first search frame vector, and second template frame vector in the first concatenation vector to obtain first normalized feature information; and then performing a residual connection between the first normalized feature information and the second attention feature information using the general residual connection method to obtain second normalized feature information. The second normalized feature information obtained from the last operation is determined as the final mixed feature information, yielding the first mixed feature information. Here, the value of n can be determined by user input or by system default. Using the same deep encoding fusion method as the target first concatenated vector, all the first concatenated vectors in the first concatenated vector set are deeply encoded and fused to obtain the first mixed feature information set.

[0051] In this example, by using an improved template matching algorithm to concatenate the first image block vectors in the set of m first image block vectors, a first concatenated vector set is obtained. The first concatenated vector in the first concatenated vector set is then input into a self-attention coding model for deep coding fusion to obtain a first mixed feature information set. This allows the first mixed feature information to reflect the dynamic change trend of the reference tracking target, thereby improving the accuracy of the UAV in tracking the target.

[0052] In one possible implementation, a method for enhancing the first feature information in the first feature information set to obtain a second feature information set includes: C1. Construct the temporal correlation interaction matrix corresponding to each first feature in the first feature information set to obtain the first interaction matrix set; C2. Perform element-wise dot product fusion on the first interaction matrix in the first interaction matrix set to obtain the first feature tensor information set. C3. Adjust the shape of the first feature tensor information according to the batch size, channel dimension, number of queries and number of spatial locations of the first feature tensor information to obtain the second feature tensor information; C4. Reconstruct the second feature tensor information into a spatial feature map structure to obtain the second feature information.

[0053] Specifically, the process can involve dividing the fused feature information corresponding to the second template frame in the first mixed feature information set to obtain the first template frame fused feature information set; performing matrix multiplication on the first feature information in the first feature information set and the first template frame fused feature information in the first template frame fused feature information set to construct the first feature information set and the temporal correlation interaction matrix corresponding to the first feature information, thereby obtaining the first interaction matrix set.

[0054] After obtaining the first set of interaction matrices, the first feature information in the first set of feature information can be fused with the first interaction matrix in the first set of interaction matrixes through element-wise dot product to generate the cross-temporal context-enhanced feature tensor corresponding to the first feature information, thereby obtaining the first feature tensor information set.

[0055] After obtaining the first set of feature tensor information, a first set of sequence length information can be obtained by acquiring the sequence length corresponding to the first feature tensor information in the first set of feature tensor information. Then, based on the first sequence length information in the first set of sequence length information, the one-dimensional first feature tensor information in the first set of feature tensor information is reconstructed into two-dimensional second feature tensor information, resulting in a second set of feature tensor information. The second feature tensor information includes the width, height, batch size, and channel dimension information of the second feature tensor. The first sequence length information and the width and height information in the second feature tensor information satisfy the following relationship: In the formula This represents the length information of the first sequence in the first sequence length information set; This represents the height information in the second feature tensor. This represents the width information in the second feature tensor.

[0056] After obtaining the second feature tensor information set, the second feature tensor information in the second feature tensor information set can be reconstructed into a spatial feature graph structure through a general tensor structure reconstruction method to obtain the second feature information set.

[0057] In this example, by constructing the temporal correlation interaction matrix corresponding to the first feature information in the first feature information set and performing element-wise weighted fusion, a first feature tensor set is obtained; the dimension of the first feature tensor information in the first feature tensor information set is adjusted and reconstructed into a spatial feature map structure to obtain a second feature information set, thereby improving the accuracy of UAV in tracking the target to be tracked.

[0058] In one possible implementation, a method for adjusting an initial tracking model based on second feature information in the second feature information set to obtain a target tracking model includes: D1. Extract the center score map of each second feature information in the second feature information set to obtain the first center score map set; D2. Extract the size prediction map of each second feature information in the second feature information set to obtain the first size prediction map set; D3. Extract the center offset map of each second feature information in the second feature information set to obtain the first center offset map set; D4. Based on the first center score map in the first center score map set, the first size prediction map in the first size prediction map set, and the first center offset map in the first center offset map set, predict the tracking box of the reference tracking target to obtain the target tracking box information set. D5. Calculate the model loss value of the initial tracking model based on the target tracking boxes in the target tracking box information set to obtain the first loss value set; D6. Determine whether all the first loss values ​​in the first loss value set are greater than the preset loss value threshold, and obtain the determination result; D7. Based on the judgment result, the model parameters of the initial tracking model are adjusted using the backpropagation algorithm until the model loss value of the adjusted tracking model is less than or equal to the preset loss value threshold, thus obtaining the target tracking model.

[0059] Specifically, the target second feature information can be obtained by arbitrarily selecting one second feature information from the second feature information set; the target second feature information is then input into a lightweight multi-branch convolutional prediction network, and the center position prediction branch of the lightweight multi-branch convolutional prediction network is used to obtain the spatial position distribution of the reference tracking target in the second search frame through a five-layer progressively decreasing channel convolutional structure, thus obtaining a first center score map. The same method used to obtain the center score map of the target second feature information is then used to obtain the center score maps corresponding to all the second feature information in the second feature information set, thus obtaining a first center score map set.

[0060] While obtaining the first center score map, the target's second feature information can be input into a lightweight multi-branch convolutional prediction network. The size regression branch of this network can then be used to estimate the size of the reference tracking target in the second search frame point-by-point, resulting in a first size prediction map. The same method used to obtain the size prediction map of the target's second feature information is then used to obtain size prediction maps corresponding to all the second feature information in the second feature information set, resulting in a first size prediction map set.

[0061] While obtaining the first size prediction map, the target's second feature information can be input into a lightweight multi-branch convolutional prediction network. The offset branch of this network can be used to estimate the adjustment information of the center point position of the reference tracking target in the second search frame, resulting in a first center offset map. The same method used to obtain the center offset map of the target's second feature information is then used to obtain the center offset maps corresponding to all the second feature information in the second feature information set, thus obtaining a first center offset map set.

[0062] After obtaining the first center score map set, the first size prediction map set, and the first center offset map set, a first position information set can be obtained by acquiring the position information of the maximum response position in each of the first center score maps in the first center score map set; a first size information set can be obtained by acquiring the size information of the reference tracking target in the second search frame of the second search frame set when the reference tracking target appears in the first position information based on the first position information in the first size prediction map of the first size prediction map set; a first size information set can be obtained by acquiring the adjustment information of the center point position of the reference tracking target in the second search frame of the second search frame set when the reference tracking target appears in the first position information based on the first position information in the first center offset map of the first center offset map set; and a target tracking box information set can be obtained by using a general decoding function to calculate the center position coordinates and width and height values ​​of the reference tracking target in the second search frame of the second search frame set based on the first position information in the first position information set, the first size information in the first size information set, and the first adjustment information in the first adjustment information set.

[0063] After obtaining the target tracking box information set, a reference label information set can be obtained by acquiring the label information corresponding to each video frame in the first video; the tracking box information corresponding to the reference tracking target in the reference label information set can be extracted to obtain a reference tracking box information set; the loss value of the initial tracking model can be calculated based on the reference tracking box information in the reference tracking box information set and the target tracking box information in the target tracking box information set to obtain a first loss value set.

[0064] Specifically, the loss value of the initial tracking model can be calculated based on the reference tracking box information in the reference tracking box information set and the target tracking box information in the target tracking box information set, as shown in the following formula, to obtain the first loss value set: In the formula This represents the first loss value in the first set of loss values; The classification loss value of the initial tracking model can be calculated using a common classification loss calculation method. The cross-union ratio loss value of the initial tracking model can be obtained by calculating the spatial overlap ratio between the tracking boxes corresponding to the reference tracking box information in the reference tracking box information set and the target tracking box information in the target tracking box information set, respectively. The weight information corresponding to the crossover ratio loss value of the initial tracking model can be determined by user input or by system default. The first-order absolute error loss value corresponding to the initial tracking model can be obtained by calculating the average value of the absolute distance difference between the target tracking box information and the reference tracking box information in each coordinate dimension. The weight information corresponding to the first-order absolute error loss value of the initial tracking model can be determined by user input or by system default. The view-invariant representation loss value of the initial tracking model can be obtained by acquiring the mutual information between the feature representations of the reference tracking target from different viewpoints; The weight information corresponding to the view invariance characterization loss value of the initial tracking model can be determined by user input or by system default. This represents the true label corresponding to the sample input during the initial training of the tracking model. When, it indicates that the input sample is a positive sample, that is, an image containing the reference tracking target, point When , it indicates that the input sample is a negative sample, that is, an image that does not contain the reference tracking target; This represents the probability that the initial tracking model can track the reference target when it receives an input sample; This indicates the location coordinates of the reference tracked target within the region corresponding to the target tracking box information in the target tracking box information set; The location coordinates of the reference tracking target within the region corresponding to the reference tracking box information in the reference tracking box information set; This represents the x-coordinate of the center point of the target tracking box within the target tracking box information set. This represents the x-coordinate of the center point of the reference tracking box within the reference tracking box information set. This represents the ordinate of the center point of the target tracking box within the target tracking box information set. This represents the ordinate of the center point of the reference tracking box within the reference tracking box information set. This indicates the width information of the target tracking box within the target tracking box information set; This indicates the width information of the reference tracking box within the reference tracking box information set; This represents the height information of the target tracking bounding boxes in the target tracking bounding box information set with respect to the target tracking bounding box. This represents the height information of the reference tracking box within the reference tracking box information set; This represents the mutual information approximation based on the Jensen–Shannon estimator in the initial tracking model; This indicates the search frame feature information within the first mixed feature information; This represents the template frame feature information in the first mixed feature information.

[0065] After obtaining the first set of loss values, it can be determined whether all the first loss values ​​in the first set are greater than a preset loss value threshold. If there is a first loss value in the first set that is greater than the preset loss value threshold, it indicates that there is a significant difference between the reference tracking box information in the reference tracking box information set and the target tracking box information set. The judgment result could be, for example, "The tracking accuracy of the current model is insufficient, and the initial tracking model needs to be adjusted." If all the first loss values ​​in the first set are less than or equal to the preset loss value threshold, it indicates that the reference tracking box information in the reference tracking box information set and the target tracking box information in the target tracking box information set have a high similarity. The judgment result could be, for example, "The tracking accuracy of the current model is high, and there is no need to adjust the initial tracking model." The above judgment results are only illustrative and do not limit the specific content of the judgment results.

[0066] After obtaining the judgment result, if the judgment result indicates that no adjustment is needed to the initial tracking model, the initial tracking model is output to obtain the target tracking model; if the judgment result indicates that the initial tracking model needs to be adjusted, the initial tracking model can be adjusted using a general AdamW (Adaptive Moment Estimation-Weight Decay) optimizer according to a general backpropagation algorithm to obtain an adjusted tracking model. The loss value corresponding to the adjusted tracking model is recalculated based on the reference tracking box information in the reference tracking box information set and the target tracking box information in the target tracking box information set to obtain a second loss value set. It is determined whether the second loss value in the second loss value set is greater than a preset loss value threshold. If there is a second loss value in the second loss value set that is greater than the preset loss value threshold, the adjusted tracking model is further adjusted using the AdamW optimizer according to a general backpropagation algorithm, and the second loss value in the second loss value set is recalculated until the second loss value in the second loss value set is less than or equal to the preset loss value threshold. The adjusted tracking model after the last adjustment is output to obtain the target tracking model.

[0067] In this example, a lightweight multi-branch convolutional prediction network is used to determine the position of the reference tracking target in the second search frame of the second search frame set, thereby obtaining a target tracking box information set. The loss value of the initial tracking model is calculated using the target tracking box information in the target tracking box information set and the reference tracking box information in the reference tracking box information set, resulting in a first loss value set. Based on the first loss value in the first loss value set, the initial tracking model is adjusted using the AdamW optimization algorithm according to the backpropagation algorithm to obtain the target tracking model, thus improving the accuracy of the UAV in tracking the target.

[0068] For examples consistent with the above embodiments, please refer to... Figure 3 , Figure 3 This is a schematic diagram of the structure of a terminal provided in an embodiment of this application, such as... Figure 3 As shown, it includes a processor, an input device, an output device, and a memory, which are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions. The program includes instructions for performing the following steps. Extract a first template frame from the first video, where the first template frame is the first video frame in the first video from the perspective of the target drone; and extract a first search frame from the video frames to be processed to obtain a first search frame set, where the video frames to be processed are the video frames in the first video other than the first template frame. The first template frame and the first search frame in the first search frame set are fused to obtain the first feature information set. The first feature information in the first feature information set is enhanced to obtain the second feature information set; The initial tracking model is adjusted based on the second feature information in the second feature information set to obtain the target tracking model; The target tracking model is used to track the target and the target tracking result is obtained.

[0069] The above mainly describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the terminal includes the corresponding hardware structure and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0070] This application embodiment can divide the terminal into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0071] For those consistent with the above, please refer to Figure 4 , Figure 4 This application provides a schematic diagram of a lightweight single-target tracking system for unmanned aerial vehicles (UAVs). Figure 4 As shown, the system includes: Image processing unit 301 is used to extract a first template frame from a first video, the first template frame being the first video frame in the first video from the perspective of the target drone, and to extract a first search frame from video frames to be processed to obtain a first search frame set, wherein the video frames to be processed are video frames in the first video other than the first template frame. Extraction unit 302 is used to perform feature fusion on the first template frame and the first search frame in the first search frame set to obtain a first feature information set; Enhancement unit 303 is used to enhance the first feature information in the first feature information set to obtain a second feature information set; The adjustment unit 304 is used to adjust the initial tracking model according to the second feature information in the second feature information set to obtain the target tracking model; The tracking unit 305 is used to track the target to be tracked using the target tracking model and obtain the target tracking result.

[0072] In one possible implementation, the extraction unit 302 is specifically used for: The first template frame is subjected to size normalization processing to obtain the second template frame; The first search frame in the first search frame set is subjected to size normalization to obtain the second search frame set; The second search frame in the second template frame and the second search frame set are combined into a dual-frame image to obtain the first dual-frame image set; The first two-frame images in the first two-frame image set are subjected to deep encoding fusion to obtain the first mixed feature information set; The fused features corresponding to the second search frame in the second search frame set in the first mixed feature information are segmented to obtain the first feature information set.

[0073] In one possible implementation, the extraction unit 302 is specifically used for: The first two-frame images in the first two-frame image set are segmented using preset segmentation pixel blocks to obtain m sets of first image blocks; where m represents the number of first two-frame images in the first two-frame image set. Linear projection is performed on the first image block in each set of first image blocks to obtain m sets of first image block vectors; The first image block vectors in the set of m first image block vectors are concatenated to obtain a first concatenated vector set. The first concatenated vector in the first concatenated vector set is input into the self-attention coding model for deep coding fusion to obtain the first mixed feature information set.

[0074] In one possible implementation, the enhancement unit 303 is specifically used for: Construct the temporal correlation interaction matrix corresponding to each first feature in the first feature information set to obtain the first interaction matrix set; The first interaction matrix in the first interaction matrix set is weighted and fused element by element to obtain the first feature tensor information set. The first feature tensor information in the first feature tensor information set is subjected to dimensionality increase processing to obtain the second feature tensor information set; The second feature tensor information in the second feature tensor information set is reconstructed into a spatial feature map structure to obtain the second feature information set.

[0075] In one possible implementation, the adjustment unit 304 is specifically used for: Extract the center score map of each second feature information in the second feature information set to obtain the first center score map set; Extract the size prediction map of each second feature information in the second feature information set to obtain the first size prediction map set; Extract the center offset map of each second feature information in the second feature information set to obtain the first center offset map set; The tracking box of the reference tracking target is predicted based on the first center score map in the first center score map set, the first size prediction map in the first size prediction map set, and the first center offset map in the first center offset map set, to obtain a target tracking box information set; The model loss value of the initial tracking model is calculated based on the target tracking boxes in the target tracking box information set to obtain the first loss value set; Determine whether all the first loss values ​​in the first loss value set are greater than a preset loss value threshold, and obtain the determination result; Based on the judgment results, the model parameters of the initial tracking model are adjusted using the backpropagation algorithm until the model loss value of the adjusted tracking model is less than or equal to the preset loss value threshold, thus obtaining the target tracking model.

[0076] This application also provides a computer storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the lightweight single-target tracking methods for unmanned aerial vehicles described in the above method embodiments.

[0077] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program that causes a computer to perform some or all of the steps of any of the lightweight single-target tracking methods for unmanned aerial vehicles described in the above method embodiments.

[0078] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0079] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0080] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0081] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] Furthermore, the functional units in the various embodiments of the application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software program module.

[0083] If the integrated unit is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0084] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory, a random access memory, a magnetic disk, or an optical disk, etc.

[0085] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A lightweight single-target tracking method for unmanned aerial vehicles (UAVs), characterized in that, The method includes: Extract a first template frame from the first video, where the first template frame is the first video frame in the first video from the perspective of the target drone; and extract a first search frame from the video frames to be processed to obtain a first search frame set, where the video frames to be processed are the video frames in the first video other than the first template frame. The first template frame and the first search frame in the first search frame set are fused to obtain the first feature information set. The first feature information in the first feature information set is enhanced to obtain the second feature information set; The initial tracking model is adjusted based on the second feature information in the second feature information set to obtain the target tracking model; The target tracking model is used to track the target and the target tracking result is obtained.

2. The lightweight single-target tracking method for unmanned aerial vehicles according to claim 1, characterized in that, The step of fusing features between the first template frame and the first search frame in the first search frame set to obtain first feature information includes: The first template frame is subjected to size normalization processing to obtain the second template frame; The first search frame in the first search frame set is subjected to size normalization to obtain the second search frame set; The second search frame in the second template frame and the second search frame set are combined into a dual-frame image to obtain the first dual-frame image set; The first two-frame images in the first two-frame image set are subjected to deep encoding fusion to obtain the first mixed feature information set; The fused features corresponding to the second search frame in the second search frame set in the first mixed feature information are segmented to obtain the first feature information set.

3. The lightweight single-target tracking method for unmanned aerial vehicles according to claim 2, characterized in that, The step of performing deep coding fusion on the first two-frame images in the first two-frame image set to obtain the first mixed feature information set includes: The first two-frame images in the first two-frame image set are segmented using preset segmentation pixel blocks to obtain m sets of first image blocks; where m represents the number of first two-frame images in the first two-frame image set. Linear projection is performed on the first image block in each set of first image blocks to obtain m sets of first image block vectors; The first image block vectors in the set of m first image block vectors are concatenated to obtain a first concatenated vector set. The first concatenated vector in the first concatenated vector set is input into the self-attention coding model for deep coding fusion to obtain the first mixed feature information set.

4. The lightweight single-target tracking method for unmanned aerial vehicles according to claim 3, characterized in that, The step of enhancing the first feature information in the first feature information set to obtain the second feature information set includes: Construct the temporal correlation interaction matrix corresponding to each first feature in the first feature information set to obtain the first interaction matrix set; The first interaction matrix in the first interaction matrix set is weighted and fused element by element to obtain the first feature tensor information set. The first feature tensor information in the first feature tensor information set is subjected to dimensionality increase processing to obtain the second feature tensor information set; The second feature tensor information in the second feature tensor information set is reconstructed into a spatial feature map structure to obtain the second feature information set.

5. The lightweight single-target tracking method for unmanned aerial vehicles according to claim 4, characterized in that, The step of adjusting the initial tracking model based on the second feature information in the second feature information set to obtain the target tracking model includes: Extract the center score map of each second feature information in the second feature information set to obtain the first center score map set; Extract the size prediction map of each second feature information in the second feature information set to obtain the first size prediction map set; Extract the center offset map of each second feature information in the second feature information set to obtain the first center offset map set; The tracking box of the reference tracking target is predicted based on the first center score map in the first center score map set, the first size prediction map in the first size prediction map set, and the first center offset map in the first center offset map set, to obtain a target tracking box information set; The model loss value of the initial tracking model is calculated based on the target tracking boxes in the target tracking box information set to obtain the first loss value set; Determine whether all the first loss values ​​in the first loss value set are greater than a preset loss value threshold, and obtain the determination result; Based on the judgment results, the model parameters of the initial tracking model are adjusted using the backpropagation algorithm until the model loss value of the adjusted tracking model is less than or equal to the preset loss value threshold, thus obtaining the target tracking model.

6. A lightweight single-target tracking system for unmanned aerial vehicles (UAVs), characterized in that, The system includes: The image processing unit is used to extract a first template frame from a first video, the first template frame being the first video frame in the first video from the perspective of the target drone, and to extract a first search frame from the video frames to be processed to obtain a first search frame set, wherein the video frames to be processed are video frames in the first video other than the first template frame. The extraction unit is used to perform feature fusion on the first template frame and the first search frame in the first search frame set to obtain a first feature information set. An enhancement unit is used to enhance the first feature information in the first feature information set to obtain a second feature information set. The adjustment unit is used to adjust the initial tracking model according to the second feature information in the second feature information set to obtain the target tracking model; The tracking unit is used to track the target to be tracked using the target tracking model and obtain the target tracking result.

7. The lightweight single-target tracking system for unmanned aerial vehicles according to claim 6, characterized in that, In terms of performing feature fusion on the first template frame and the first search frame in the first search frame set to obtain first feature information, the extraction unit is specifically used for: The first template frame is subjected to size normalization processing to obtain the second template frame; The first search frame in the first search frame set is subjected to size normalization to obtain the second search frame set; The second search frame in the second template frame and the second search frame set are combined into a dual-frame image to obtain the first dual-frame image set; The first two-frame images in the first two-frame image set are subjected to deep encoding fusion to obtain the first mixed feature information set; The fused features corresponding to the second search frame in the second search frame set in the first mixed feature information are segmented to obtain the first feature information set.

8. The lightweight single-target tracking system for unmanned aerial vehicles according to claim 7, characterized in that, In terms of performing deep coding fusion on the first two-frame images in the first two-frame image set to obtain the first mixed feature information set, the extraction unit is specifically used for: The first two-frame images in the first two-frame image set are segmented using preset segmentation pixel blocks to obtain m sets of first image blocks; where m represents the number of first two-frame images in the first two-frame image set. Linear projection is performed on the first image block in each set of first image blocks to obtain m sets of first image block vectors; The first image block vectors in the set of m first image block vectors are concatenated to obtain a first concatenated vector set. The first concatenated vector in the first concatenated vector set is input into the self-attention coding model for deep coding fusion to obtain the first mixed feature information set.

9. A terminal, characterized in that, The device includes a processor, an input device, an output device, and a memory, which are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the lightweight single-target tracking method for unmanned aerial vehicles as described in any one of claims 1-5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the lightweight single-target tracking method for unmanned aerial vehicles as described in any one of claims 1-5.

Images