An unmanned aerial vehicle tracking method and system based on a GEM-Track tracker

Abstract

The application discloses a kind of unmanned aerial vehicle tracking method and system based on GEM-Track tracker, the method includes: obtaining unmanned aerial vehicle video sequence and respectively carries out initialization and frame-by-frame tracking processing, obtains template tensor and search tensor;Based on SGLATrack single-branch visual Transformer tracking architecture, introduce multi-scale positioning enhancement auxiliary prediction head and metric perception optimization module, construct GEM-Track unmanned aerial vehicle tracker;Based on GEM-Track unmanned aerial vehicle tracker, template tensor and search tensor are tracked, realize the real-time tracking of unmanned aerial vehicle.The application can improve unmanned aerial vehicle tracking efficiency and accuracy, and reduce the additional calculation cost of tracking.The application as a kind of unmanned aerial vehicle tracking method and system based on GEM-Track tracker, can be widely applied in unmanned aerial vehicle tracking technical field.

Description

Technical Field

[0001] This invention relates to the field of drone tracking technology, and in particular to a drone tracking method and system based on a GEM-Track tracker. Background Technology

[0002] Unmanned aerial vehicle (UAV) tracking has become a key research area, with applications spanning security patrols, path planning, border surveillance, and precision agriculture. However, compared to general target tracking, UAV tracking faces more severe challenges due to its unique aerial perspective: drastic changes in target scale (target pixel size can differ by 5-10 times depending on flight altitude), frequent occlusion, complex backgrounds, and the tendency for target apparent drift during long-term tracking. More importantly, the limited endurance and computing resources of UAVs place stringent demands on tracker efficiency, requiring real-time performance (≥30 frames / second).

[0003] Thanks to the efficiency of Fourier domain operations, discriminant correlation filtering (DCF)-based methods are widely used in UAV tracking tasks. Despite their high efficiency, these methods rely on hand-designed feature extractors, making it difficult to achieve ideal tracking robustness. Siamese network-based trackers can learn robust target representations through deep feature learning, but their accuracy still has significant room for improvement due to their reliance on early convolutional neural networks (CNNs). In recent years, visual Transformer (ViT)-based trackers have received increasing attention. Typical examples include MixFormer and OSTrack, which have driven the evolution of trackers towards highly parallel single-branch architectures. However, their computational cost remains high, making direct deployment in real-time UAV tracking scenarios difficult. To improve efficiency, AbaViTrack introduces an adaptive background-aware token termination strategy in miniature ViT; also based on miniature ViT, AVTrack proposes a more structured method that dynamically enables or disables ViT layers based on input complexity. While these works validate the application potential of miniature ViT, they primarily focus on optimizing the efficiency of the backbone network, and the core challenges in prediction head design and template management remain largely unresolved. More fundamentally, existing methods have the following shortcomings: 1) The dilemma of prediction heads: complex prediction heads have high computational costs, while simple single-scale prediction heads are difficult to cope with multi-scale targets and background interference; 2) Template drift: The performance of a fixed template degrades when the target appearance changes, while blindly updating it introduces noise and amplifies the accumulated error in long-term tracking. 3) Orthogonal optimization gap: Most works only optimize the efficiency of the backbone network, ignoring the synergistic potential to simultaneously enhance spatial positioning accuracy and temporal template consistency. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a UAV tracking method and system based on a GEM-Track tracker, which can improve the efficiency and accuracy of UAV tracking and reduce the additional computational costs of tracking.

[0005] The first technical solution adopted in this invention is: a drone tracking method based on a GEM-Track tracker, comprising the following steps: The drone video sequence is acquired and initialized and tracked frame by frame to obtain the template tensor and the search tensor. Based on the SGLATrack single-branch visual Transformer tracking architecture, a multi-scale localization enhancement auxiliary prediction head and a metric perception optimization module are introduced to construct the GEM-Track UAV tracker. Based on the GEM-Track drone tracker, the template tensor and search tensor are tracked to achieve real-time tracking of the drone.

[0006] Furthermore, the step of acquiring the UAV video sequence and performing initialization and frame-by-frame tracking processing to obtain the template tensor and search tensor specifically includes: Acquire drone video sequences and extract the first frame image and its labeled target bounding box; Given a clipping factor, determine the clipping region using the target bounding box and calculate the side length of the clipping region; The first frame image is cropped according to the side length of the cropping region to obtain the template image; The template image is preprocessed using standardization to obtain the template tensor; Based on the second frame of the drone video sequence, frame-by-frame tracking is performed to obtain the original image of the current frame; Based on the target bounding box of the previous frame, a basic search cropping factor is set and adaptively adjusted. The original image of the current frame is cropped to obtain the search image. After standardization preprocessing, the search tensor is obtained.

[0007] Furthermore, the GEM-Track UAV tracker specifically includes an initial template buffer, an online template buffer, a patch embedding layer, a ViT backbone network, a main prediction head, a multi-scale localization enhancement auxiliary prediction head, and a metric perception optimization module.

[0008] Furthermore, the step of tracking the template tensor and search tensor based on the GEM-Track drone tracker to achieve real-time tracking of the drone specifically includes: Input the template tensor and the search tensor into the GEM-Track drone tracker; An initial template buffer and an online template buffer based on the GEM-Track UAV tracker, wherein the initial template buffer and the online template buffer are used to store template tensors; Based on the patch embedding layer of the GEM-Track UAV tracker, the template tensor and search tensor of the online template buffer are divided into non-overlapping image blocks and mapped into feature vectors to obtain several template tags and several search tags. Learnable positional codes are added and concatenated to obtain a concatenated tag sequence. Based on the ViT backbone network of the GEM-Track UAV tracker, the search region features are separated and reshaped after ViT processing of the spliced ​​labeled sequence to obtain deep semantic features. The deep semantic features are first expanded in size using adaptive average pooling, and then restored in size using bilinear interpolation to obtain shallow features. Based on the main prediction head of the GEM-Track UAV tracker, multi-layer convolution processing is performed on deep semantic features to obtain center point score map, target size prediction map and sub-pixel offset prediction map. A multi-scale localization enhancement auxiliary prediction head based on the GEM-Track UAV tracker is used to perform auxiliary prediction on deep semantic features and shallow features to obtain an auxiliary confidence map. By combining the auxiliary confidence map, center point score map, target size prediction map and subpixel offset prediction map, score map fusion and bounding box decoding are performed to obtain the final fused score map, response map and target bounding box prediction result of the current frame; The metric perception optimization module based on the GEM-Track UAV tracker performs metric extraction and template fusion on the final fused score map and response map to obtain the UAV tracking results.

[0009] Furthermore, the step of using the multi-scale localization enhancement auxiliary prediction head based on the GEM-Track UAV tracker to perform auxiliary prediction on deep semantic features and shallow features to obtain an auxiliary confidence map specifically includes: Deep semantic features and shallow features are input into the multi-scale localization enhancement auxiliary prediction head of the GEM-Track UAV tracker; The deep semantic features are processed by generating main features and auxiliary features respectively. The main features and auxiliary features are output and concatenated to obtain the aligned deep semantic features. Align the shallow features with channels and resolution to obtain aligned shallow features; Based on the global gating mechanism, the aligned deep semantic features and the aligned shallow features are combined into fusion candidate features. Dynamic weights are determined, and scalar gating weights and weighted deep weights are generated according to the dynamic weights. The aligned deep semantic features and the aligned shallow features are fused at multiple scales to obtain multi-scale gating fusion features. A channel recalibration mechanism is introduced to generate channel attention weights to recalibrate the multi-scale gated fusion features, thereby obtaining an auxiliary confidence map.

[0010] Furthermore, the step of combining the auxiliary confidence map, center point score map, target size prediction map, and sub-pixel offset prediction map to perform score map fusion and bounding box decoding, and obtaining the final fused score map, response map, and target bounding box prediction result for the current frame, specifically includes: The auxiliary confidence map and the center point score map are fused with equal weights to obtain the final fused score map; A Hanning window is applied to the final fused score map to suppress edge responses, resulting in a response map. Based on the response map, the peak position of the response map is obtained. Combined with the target size prediction map and the sub-pixel offset prediction map, the normalized bounding box coordinates of the target are decoded. The occlusion count results from the metric-aware optimization module are used to adaptively adjust the basic search pruning factor. The target normalized bounding box coordinates are then combined to perform coordinate mapping, resulting in the target bounding box prediction for the current frame.

[0011] Furthermore, the metric perception optimization module based on the GEM-Track UAV tracker performs metric extraction and template fusion on the final fused score map and response map to obtain the UAV tracking result. This step specifically includes: The final fused score map and response map are input into the metric perception optimization module of the GEM-Track drone tracker; Based on the quality metric extraction unit of the metric perception optimization module, the final fused score map and response map are metric extracted to obtain ternary metric information, which includes response entropy, kurtosis and confidence. An occlusion status discrimination mechanism and update conditions are set based on ternary metric information; If the update conditions are met, a new template image block is cropped based on the target bounding box prediction result of the current frame as a candidate template. The candidate template is then fused with the online template maintained in the previous time step using the exponential moving average algorithm to obtain the fused template image. Based on the occlusion counting results of the metric perception optimization module, the first frame image and the fused template image are introduced for anchor fusion to obtain a new fused template image, which is written back to the online template buffer. The target bounding box coordinates of each frame are output as the UAV tracking result.

[0012] The second technical solution adopted in this invention is: a drone tracking system based on a GEM-Track tracker, comprising: The first module is used to acquire UAV video sequences and perform initialization and frame-by-frame tracking processing to obtain template tensors and search tensors. The second module is used to build the GEM-Track UAV tracker based on the SGLATrack single-branch visual Transformer tracking architecture by introducing a multi-scale localization enhancement auxiliary prediction head and a metric perception optimization module. The third module is used to track template tensors and search tensors based on the GEM-Track drone tracker, thereby enabling real-time tracking of drones.

[0013] The beneficial effects of the method and system of this invention are as follows: This invention acquires UAV video sequences and performs initialization and frame-by-frame tracking processing to obtain template tensors and search tensors; further, based on the SGLATrack single-branch visual Transformer tracking architecture, it introduces a multi-scale localization enhancement auxiliary prediction head and a metric-aware optimization module to construct the GEM-Track UAV tracker. The multi-scale localization enhancement auxiliary prediction head uses Ghost convolution and adaptive gating mechanisms to achieve lightweight multi-scale feature fusion, effectively enhancing the tracker's robustness to target scale changes. The metric-aware optimization module performs quality-aware template updates under zero-parameter conditions through occlusion discrimination of ternary metrics and adaptive dynamic EMA fusion of the dataset, which helps reduce the cumulative error in long-term tracking; finally, based on the GEM-Track UAV tracker, the template tensor and search tensor are tracked to achieve real-time tracking of the UAV, which can improve the efficiency and accuracy of UAV tracking and reduce the additional computational cost of tracking. Attached Figure Description

[0014] Figure 1 This is a flowchart of the steps of a UAV tracking method based on a GEM-Track tracker according to the present invention; Figure 2 This is a structural block diagram of a UAV tracking system based on a GEM-Track tracker according to the present invention; Figure 3 This is a schematic diagram illustrating the accuracy-speed comparison on the UAVDT dataset provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the GET-Head structure provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the overall architecture of the GEM-Track tracker provided in a specific embodiment of the present invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0016] First, it's important to note that lightweight visual Transformers (ViTs) have demonstrated exceptional efficiency in real-time UAV tracking. However, existing single-branch trackers suffer from significant limitations: the prediction head struggles to balance multi-scale feature aggregation with computational efficiency, while template management strategies are prone to target appearance drift or noise accumulation. This study finds that implementing orthogonal optimization in the spatial and temporal dimensions can significantly improve the performance of lightweight ViT trackers with only a lower computational overhead. Based on this finding, we propose the GEM-Track tracker (a UAV tracker based on a ghost-enhanced head and metric-aware template optimization), which integrates two plug-and-play modules into an existing high-efficiency tracker architecture. The Ghost-enhanced multi-scale tracking head (GET-Head) leverages Ghost convolution and adaptive gating mechanisms to achieve lightweight multi-scale feature fusion, effectively enhancing the tracker's robustness to target scale changes. Simultaneously, the metric-aware template optimization module (MATR), based on a ternary metric discrimination mechanism and a quality-aware dynamic fusion strategy, achieves template optimization without additional training, effectively suppressing accumulated drift in long-term tracking scenarios. Multiple experiments conducted on five highly challenging drone tracking benchmark datasets validated the effectiveness of our method: significant performance improvements can be achieved by introducing only a small amount of additional computational cost while maintaining real-time tracking efficiency.

[0017] Reference Figure 1 This invention provides a drone tracking method based on a GEM-Track tracker, the method comprising the following steps: S100: Acquire the UAV video sequence and perform initialization and frame-by-frame tracking processing respectively to obtain the template tensor and the search tensor; Specifically, the process involves acquiring a UAV video sequence, extracting the first frame image and its labeled target bounding box, determining the cropping region based on the target bounding box, and calculating the side length of the cropping region, cropping the first frame image according to the side length of the cropping region to obtain a template image, performing standardization preprocessing on the template image to obtain a template tensor, and performing frame-by-frame tracking based on the second frame of the UAV video sequence to acquire the original image of the current frame, setting a basic search cropping factor based on the target bounding box of the previous frame and adaptively adjusting it, cropping the original image of the current frame to obtain a search image, performing standardization preprocessing to obtain a search tensor.

[0018] S200, based on the SGLATrack single-branch visual Transformer tracking architecture, introduces a multi-scale localization enhancement auxiliary prediction head and a metric perception optimization module to build the GEM-Track UAV tracker; Specifically, the GEM-Track UAV tracker includes an initial template buffer, an online template buffer, a patch embedding layer, a ViT backbone network, a main prediction head, a multi-scale localization enhancement auxiliary prediction head, and a metric perception optimization module.

[0019] The S300, based on the GEM-Track drone tracker, tracks template tensors and search tensors to achieve real-time tracking of drones.

[0020] In this embodiment, the present invention integrates the GET-Head and MATR modules into the SGLATrack single-branch visual Transformer tracking architecture to construct a high-efficiency tracker for real-time UAV tracking, named GEM-Track. To clearly demonstrate the versatility of GEM-Track's design, the overall architecture uses a standard single-branch visual Transformer as the backbone network, emphasizing the plug-and-play nature and universal adaptability of the two modules, such as... Figure 3 The diagram shows the overall architecture of the GEM-Track tracker. This framework integrates the GET-Head and MATR modules into a single-branch ViT tracking architecture: after the template image and search image are processed by patch embedding and the ViT backbone network, the GET-Head generates an auxiliary confidence map and fuses it with the main prediction head result, and the MATR adaptively manages the template update based on the quality metric.

[0021] S310. Input the template tensor and the search tensor into the GEM-Track drone tracker; S320. An initial template buffer and an online template buffer based on the GEM-Track UAV tracker, wherein the initial template buffer and the online template buffer are used to store template tensors; In this embodiment, the first frame image of the video sequence and its labeled target bounding box are first obtained. Using the center of the target bounding box as a reference, and based on the template clipping factor... Calculate the side length of the cutting area A square region centered on the target is cropped from the first frame and scaled to a fixed size of 128×128 to obtain the template image. The template image is then subjected to normalization preprocessing (normalized to ImageNet mean and standard deviation) to generate a template tensor. For example... Figure 3As shown, the template tensor is stored in both the Initial Template Buffer (ITB) and the Online Template Buffer (OTB): the initial template is stored in the ITB. It remains unchanged throughout the tracking process and will not be overwritten by subsequent updates; it serves only as a long-term semantic anchor for template fusion during occlusion recovery; online templates stored in the OTB. (during initialization and) (Consistent) serves as the real-time input to the backbone network, and is subsequently dynamically updated by the MATR module. Simultaneously, the target bounding box of the first frame is recorded as an initial tracking state and a cache of valid states.

[0022] Starting from the second frame, frame-by-frame tracking is performed. First, the original image of the current frame is acquired. Using the target state (boundary box position and scale) of the previous frame as a reference, the basic search cropping factor is calculated. If the MATR module detects that the target is occluded, then press... Dynamically expand the clipping factor, where the expansion amount The occlusion count increases linearly and has an upper limit, thereby expanding the search range and increasing the probability of target re-capture.

[0023] The cropping side length is calculated based on the expanded cropping factor. A square search region, with the target center of the previous frame as the reference, is cropped from the current frame and scaled to a fixed size of 256×256 to obtain the search image. The same normalization preprocessing as the template is performed on the search image to generate a search tensor, and the scaling factor is recorded. This is used for subsequent coordinate inverse mapping.

[0024] S330, a patch embedding layer based on the GEM-Track UAV tracker, divides the template tensor and search tensor of the online template buffer into non-overlapping image blocks and maps them into feature vectors to obtain several template tags and several search tags. Learnable positional codes are added and concatenated to obtain a concatenated tag sequence. In this embodiment, the online template in the OTB is used. The search tensor and the image are fed into a shared patch embedding layer. This layer divides the image into 16×16 non-overlapping image blocks, and maps each image block to a 16×16 non-overlapping image block using a convolution with a stride equal to the block size. 3D feature vector: obtained from template image Template markup Searching for images yields search tags Subsequently, learnable positional codes were added to the template tags and search tags respectively, and then concatenated along the tag dimension to form a unified sequence. (in ), which serves as the input to the backbone network.

[0025] S340, based on the ViT backbone network of the GEM-Track UAV tracker, performs ViT processing on the spliced ​​labeled sequence to separate the search region features and reshape the features to obtain deep semantic features; In this embodiment, the spliced ​​marker sequence Input by The ViT backbone network consists of layers of Transformer blocks. Each Transformer block contains a multi-head self-attention and feedforward network, where template labels and search labels complete sufficient feature interactions and relationship modeling within the same attention space. After layer processing, the backbone network outputs a fused feature sequence that has been normalized by the layers, and the search region features are separated from it according to the labeled positions. The features corresponding to the search region are then extracted. Each marker is reshaped into a two-dimensional spatial feature map. (in , ), as a deep semantic feature.

[0026] S350. Deep semantic features are first expanded in size using adaptive average pooling, and then restored in size using bilinear interpolation to obtain shallow features. In this embodiment, to provide multi-scale input to the GET-Head, deep features are processed. First, the spatial size is expanded to [a certain value] using adaptive average pooling. Then, bilinear interpolation is used to restore it to its original size. shallow features are obtained. This zoom-in and zoom-out operation introduces a local smoothing effect, simulating the detail-perceiving ability of shallow networks. (Deep features) shallow features Together they constitute the output characteristics of the backbone network.

[0027] S360, based on the GEM-Track UAV tracker's main prediction head, performs multi-layer convolution processing on deep semantic features to obtain center point score map, target size prediction map and sub-pixel offset prediction map; In this embodiment, deep features are... The data is fed into the primary prediction head. This prediction head contains three parallel multi-layer convolutional branches: 1) Central point branch: via four layers After progressive dimensionality reduction through convolution, Convolution and Sigmoid activation output center point score map ; 2) Size Branch: Using the same structure, output the target size prediction map. ; 3) Offset Branch: Outputs sub-pixel offset prediction map .

[0028] S370, a multi-scale localization enhancement auxiliary prediction head based on GEM-Track UAV tracker, performs auxiliary prediction on deep semantic features and shallow features to obtain auxiliary confidence maps; Specifically, deep semantic features and shallow features are input into the multi-scale localization enhancement auxiliary prediction head of the GEM-Track UAV tracker; the deep semantic features are processed by generating main features and auxiliary features respectively, and the main features and auxiliary features are output and concatenated to obtain aligned deep semantic features; the shallow features are aligned by channel and resolution to obtain aligned shallow features; based on a global gating mechanism, the aligned deep semantic features and aligned shallow features are synthesized into fusion candidate features, dynamic weights are determined, and scalar gating weights and weighted deep weights are generated according to the dynamic weights. The aligned deep semantic features and aligned shallow features are fused at multiple scales to obtain multi-scale gated fusion features; a channel recalibration mechanism is introduced to generate channel attention weights to recalibrate the multi-scale gated fusion features to obtain an auxiliary confidence map.

[0029] In this embodiment, shallow features with deep features Simultaneously input the GET-Head module. GET-Head performs the following operations in sequence: performs lightweight compression and alignment of deep features using GhostModule; and... Convolutional layers perform channel alignment on shallow features and combine bilinear interpolation to achieve spatial alignment; an adaptive gating fusion mechanism is used to dynamically weight and fuse features at two scales; parallel convolutions with different dilation rates are used to extract multi-scale receptive field features; a channel recalibration mechanism is used to enhance target-related channels and suppress background noise; finally, ... Convolution and Sigmoid activation output an auxiliary confidence map. .

[0030] In some specific implementations, traditional predictive heads face a core dilemma in UAV tracking tasks: complex predictive heads generate significant computational overhead through dense feature interactions, while simple predictive heads rely on single-scale features and struggle to handle multi-scale targets. This contradiction is particularly pronounced in UAV scenarios, where target scales can vary by 5-10 times due to flight altitude. Drawing on the successful experience of GhostNet in achieving lightweight design through "master feature generation + inexpensive transformation," we designed GET-Head to simultaneously address these two challenges. This module uses the Ghost module for efficient feature compression and performs multi-scale fusion through an adaptive gating mechanism, achieving robust localization with only 0.1M parameters and 0.02G FLOPs. GET-Head works in parallel with the master predictive head, improving tracking stability by fusing their confidence maps with equal weights, as detailed below. Figure 4 As shown.

[0031] First, feature compression and alignment are performed. GhostModule achieves lightweight feature compression through a two-stage process while ensuring feature alignment. Given deep features... The target output channel is The transformation process is as follows: 1) Main Feature Generation: Extracting core information through 1×1 convolution: ; in The Ghost ratio, as verified by parameter sensitivity analysis, yields a relatively optimal balance.

[0032] 2) Low-cost auxiliary feature generation: Utilizing depthwise separable convolutions to supplement detailed information with low overhead: ; DW-Conv stands for Depthwise Separable Convolution, which decomposes standard convolution into spatial convolution and pointwise convolution to improve computational efficiency.

[0033] 3) Feature assembly: Assembling primary and secondary features: ; For shallow features Align channels with resolution: ; ; in express The spatial dimensions. After alignment, and Maintaining consistent size and resolution facilitates subsequent fusion.

[0034] Further, multi-scale gating fusion is conducted. In UAV tracking, the drastic changes in target scale make balancing shallow details and deep semantics extremely challenging. Simple fixed-weight fusion tends to favor a single scale, leading to decreased robustness. Borrowing from the attention mechanism's idea of ​​globally context-adaptive weighted features, this paper designs a global gating mechanism to dynamically balance multi-scale features. First, the aligned features are synthesized into fusion candidates. Used to calculate dynamic weights : ; ; ; GAP represents global average pooling. It is Sigmoid. Activate Hardswish , 1×1 convolution ( (Grouped convolutions are used to improve efficiency). Here... Scalar gating weights for shallow features (shared across all spatial locations and channels). Then, weighted deep features are used. Adaptive fusion is then performed as follows: ; This mechanism automatically emphasizes shallow details when dealing with small targets. Enlargement), highlighting deeper semantics in complex contexts ( (Reduce). Subsequently, to adapt to scale changes, parallel convolutions with different dilation rates are used to capture multi-scale receptive fields: ; ; ; Branch 1 ( Standard convolution focuses on capturing fine-grained edges of small targets, while branch 2 ( (Diffuse convolution) expands the global receptive field to handle large targets.

[0035] Finally, channel recalibration and target prediction are performed. The background of UAV tracking scenarios typically contains a large amount of interference information; without channel-level optimization, background noise easily dominates the prediction results. Drawing inspiration from the channel attention paradigm, we introduce a channel recalibration mechanism to enhance target-related channels and suppress background noise: ; ; ; in Extract channel-level context. The channel attention weights are calculated using a lightweight network. , (for 1×1 convolution) This indicates that element-wise multiplication is used to perform feature recalibration. A confidence map is then generated. ; Finally, the confidence plot of GET-Head and the score plot of the main predictor head are fused using equal weighting: ; in This is the score map of the original master prediction head; the size and offset are still output by the master prediction head (GET-Head only enhances the localization score). This integration method requires only 0.1M parameters and 0.02G FLOPs, effectively improving localization robustness with low overhead.

[0036] S380: Combine the auxiliary confidence map, center point score map, target size prediction map and sub-pixel offset prediction map to perform score map fusion and bounding box decoding, and obtain the final fused score map, response map and target bounding box prediction result of the current frame; Specifically, the auxiliary confidence map and the center point score map are fused with equal weights to obtain the final fused score map; a Hanning window is applied to the final fused score map to suppress edge responses, resulting in a response map; based on the response map, the peak position of the response map is obtained, and the target size prediction map and the sub-pixel offset prediction map are combined to decode the target normalized bounding box coordinates; the basic search pruning factor is adaptively adjusted based on the occlusion count results of the metric-aware optimization module, and coordinate mapping is performed in combination with the target normalized bounding box coordinates to obtain the target bounding box prediction result for the current frame.

[0037] In this embodiment, the master predictor head score map is used. With GET-Head Auxiliary Confidence Graph Merge by equal weight (e.g.) Figure 5 middle As shown), the final fusion score map is obtained: ; A Hanning window was then applied to the fused score map to suppress edge responses, resulting in a response map. Based on the peak location of the response map, combined with the size prediction map... With offset prediction map The target's normalized bounding box coordinates (center point, width, height) are decoded. Finally, the recorded scaling factor is used. The normalized coordinates are inversely mapped to the original image coordinate system and the boundaries are cropped to obtain the target bounding box prediction result for the current frame. .

[0038] Fusion Score Chart With response diagram Simultaneously serving as input to the MATR module ( Figure 5 (Note: This is marked as S / R).

[0039] S390, a measurement perception optimization module based on the GEM-Track UAV tracker, performs measurement extraction and template fusion on the final fused score map and response map to obtain the UAV tracking result.

[0040] Specifically, the final fused score map and response map are input into the metric perception optimization module of the GEM-Track UAV tracker; based on the quality metric extraction unit of the metric perception optimization module, the final fused score map and response map are metrically extracted to obtain ternary metric information, which includes response entropy, kurtosis, and confidence; based on the ternary metric information, an occlusion state discrimination mechanism and update conditions are set; if the update conditions are met, a new template image block is cropped according to the target bounding box prediction result of the current frame as a candidate template, and the candidate template is fused with the online template maintained in the previous time step using an exponential moving average algorithm to obtain a fused template image; based on the occlusion count result of the metric perception optimization module, the first frame image and the fused template image are introduced for anchor fusion to obtain a new fused template image, which is written back to the online template buffer, and the target bounding box coordinates of each frame are output as the UAV tracking result.

[0041] In this embodiment, as Figure 5 As shown in the MATR module, the metric extraction (ME) module uses the obtained fusion score map. With response diagram (Right now Figure 5 Using S / R as input, extract three complementary quality metrics: 1) From the response graph Calculate response entropy This measures the uncertainty of the response distribution. 2) From the fusion score map Calculate kurtosis This measures the significance of the peak response. 3) Extract confidence from the fusion score map .

[0042] The obtained ternary measure , , ( Figure 5 The value marked as CPE is passed to the subsequent occlusion detection module.

[0043] Occlusion detection results With confidence level Input update decision module (UD). Template update requires the following conditions to be met simultaneously: current frame number is greater than the starting frame, the number of historical updates has not reached the upper limit, the interval with the last update is not less than the minimum interval threshold, and the current frame confidence level is [not specified]. Not lower than the adaptive quality threshold And there is no occlusion in the current frame ( ).

[0044] If the update conditions are not met ( Then the online templates in OTB Remain unchanged and be used directly in the next frame (corresponding to) Figure 5 In the middle, "UD=0(Keep If the update conditions are not met for several consecutive frames, the module will gradually relax the quality threshold. With a minimum update interval, it enhances the ability to adapt to long-term appearance changes.

[0045] If the update conditions are met ( Then, using the predicted target bounding box in the current frame as the center, a new template image patch is cropped from the current frame according to the same cropping and preprocessing process to obtain the candidate template. ( Figure 5 (The text is marked as CT). With OTB Both are fed into the Template Fusion (TF) module and fused using Exponential Moving Average (EMA): ; in For EMA fusion coefficient, when kurtosis Dynamic reduction This reduces the impact of low-quality frames on updates.

[0046] In typical non-occlusion restoration scenarios, the fusion result is directly used as the updated template. When occlusion recovery is detected (occlusion counter is greater than 0), the TF module additionally reads the initial template from the ITB. Perform anchored fusion: (correspond Figure 5 The path during the "occlusion reply period" is used to ensure semantic consistency.

[0047] Received Write back to OTB replacement (Right now ← In the next frame, the template update control is sent to the backbone network via the patch embedding layer, forming a closed-loop template update control.

[0048] After the above steps, GEM-Track outputs the target bounding box coordinates for each frame. As a tracking result, the three components—main prediction head, GET-Head, and MATR—have clear interfaces and are decoupled from each other, allowing for seamless integration into any single-branch visual Transformer tracking framework without modifying the loss function or the overall network architecture.

[0049] In some specific embodiments, existing ViT trackers generally adopt the paradigm of "fixed template in the first frame + subsequent search," but this has two main drawbacks: first, template drift: the fixed template deviates from the true target distribution under changes in appearance, leading to response degradation; second, update risk: simple frequent updates introduce noise in low-confidence frames, while not updating at all cannot adapt to long-term tracking scenarios. Traditional solutions rely on manually setting thresholds or additional networks, and the introduced computational overhead conflicts with the real-time requirements of UAVs. Drawing on recent research on parameterless adaptive mechanisms and based on the observation that "tracking quality can be quantified through mathematical metrics," this paper designs the MATR module to achieve quality-aware template optimization without introducing trainable parameters. This module adopts a discrimination mechanism based on a ternary metric (response entropy, kurtosis, and confidence) and a quality-aware dynamic fusion strategy to avoid single metric bias, while having no additional computational overhead in the inference phase.

[0050] The MATR process comprises four key steps: extracting quality metrics from the response and score maps; occlusion detection based on ternary metrics; making template update decisions based on multi-condition combinations; and employing adaptive EMA fusion with occlusion recovery capabilities. The specific steps are as follows: 1) Extraction of quality metrics: MATR is based on response graphs With score chart The three complementary metrics are computed using efficient GPU tensor operations: Response entropy ( ):use A normalized histogram with 20 bars is used to assess the uncertainty of the response distribution. ; in A higher value indicates a more dispersed response distribution (unreliable tracking). The lower the value, the more concentrated the response distribution (reliable tracking).

[0051] Kubularity ( ): Assess the sharpness of the response: ; in A higher value indicates a more significant peak in the response (clearer target). This indicates that the response map is flat (due to occlusion).

[0052] Confidence level ( ): Extracted directly from the score graph: in A higher value indicates better target visibility. A lower value indicates that there may be occlusion.

[0053] 2) Occlusion status determination: To avoid misjudgments caused by a single metric, we designed a discrimination condition based on a ternary metric to accurately determine the occlusion state: in Indicate whether there is obstruction. For indicator functions, Indicates logical AND, threshold value. When the occlusion condition is met ( When occlusion occurs, the occlusion count increments, triggering an adaptive search area expansion mechanism (the search range increases linearly with the occlusion count), and the tracking state is rolled back to the most recent valid state. To handle long-term occlusion, when consecutive occluded frames exceed a preset threshold, the module will forcibly execute state rollback and reset the counter to eliminate error accumulation and enhance recovery capabilities.

[0054] 3) Template update decision: MATR determines whether to update the template based on a combination of multiple conditions. An update requires all of the following to be met simultaneously: Basic conditions: The current frame number exceeds the starting frame, the number of historical updates has not exceeded the limit, and the minimum interval since the last update has been met; Quality conditions: confidence level Not lower than the adaptive threshold ; Non-occlusion condition: No occlusion occurs in the current frame ( ).

[0055] In addition, to address the challenges of long-term tracking, the module introduces a threshold relaxation mechanism: if no update is achieved for several consecutive frames, the threshold is gradually relaxed. The update interval is adjusted to improve the algorithm's recovery chances.

[0056] 4) Template fusion strategy: When the update conditions are met, MATR crops a new template image patch from the current frame image, centered on the predicted target bounding box, as a candidate template. .set up This is the online template that was maintained in the previous step. This serves as the initial template for frame 1. MATR employs an exponential moving average (EMA) method for template fusion, with the fusion coefficients dynamically adjusted based on the response quality of the current frame. ; in This is the exponential moving average fusion coefficient. When the kurtosis P is below the sharpness threshold (P < 1.6), The value is dynamically reduced by 0.1 (with a minimum of 0.55) to further suppress the impact of low-quality frame updates. It should be noted that the fusion coefficient here... Gating weights in GET-Head These are different parameters, each acting on a different module.

[0057] Under normal (non-occlusion recovery) conditions, the online template is directly updated to... When occlusion recovery is detected (occlusion counter is greater than 0), an additional initial template is introduced for anchoring fusion to enhance semantic consistency. The specific occlusion recovery fusion expression is as follows: ; In both cases, the obtained All replaced As an online template for subsequent frames (i.e.) ← This is then sent to the patch embedding module of the next frame, forming a closed-loop template update control. It is important to emphasize that the initial template... Permanently retained throughout the entire tracking process, and will not be... Overlay; it is only used as a long-term semantic anchor point and is introduced in the occlusion recovery fusion formula to prevent cumulative drift from damaging the target identity information of the online template.

[0058] All calculations in MATR are based on efficient GPU tensor operations (histogram calculation, maximum value extraction, conditional judgment, etc.). Its additional overhead mainly comes from metric caching and conditional judgment, which is at a very low level and can meet the real-time requirements of UAV tracking. This module does not introduce any trainable parameters, and the increase in the number of floating-point operations during the inference phase is negligible.

[0059] In summary, we explore lightweight optimization strategies suitable for the core challenges of UAV tracking through complementary improvements in spatial and temporal dimensions. The design of this method is based on two observations: lightweight multi-scale feature fusion can improve localization robustness with lower computational overhead; and quality-aware template optimization can reduce drift and accumulated errors without additional trainable parameters. To this end, we propose the GEM-Track tracker, whose core consists of two plug-and-play modules that can be embedded into existing single-branch ViT models. First, GET-Head, as a lightweight auxiliary prediction head, achieves multi-scale fusion through Ghost-based feature compression (0.1M parameters, 0.02 GFLOPs), fuses shallow details and deep semantics using an adaptive gating mechanism, and captures multi-scale receptive fields through parallel convolutions with different dilation rates. GET-Head works in conjunction with the main prediction head through equal-weighted confidence map fusion to improve localization accuracy with lower overhead. Second, MATR achieves template optimization under zero-parameter conditions by fusing occlusion discrimination using ternary metrics with quality-aware dynamic exponential moving average (EMA), helping to alleviate accumulated errors in long-term tracking. By integrating these modules into a single-branch ViT, GEM-Track achieved highly competitive performance on UAV benchmark datasets: a 1.6% improvement in AUC and a 2.5% improvement in accuracy compared to suboptimal methods, while maintaining real-time speed. Figure 5 As shown, GEM-Track provides an excellent accuracy-efficiency trade-off for UAV tracking tasks. Figure 5 The image shows an accuracy-speed comparison on the UAVDT dataset. The GEM-Track tracker proposed in this embodiment achieves the highest AUC value of 61.4% among all trackers, while maintaining a competitive speed of 164.87 frames per second, validating its excellent accuracy-efficiency balance performance in real-time UAV tracking tasks. This embodiment differs from existing technologies in the following key technical features: 1) GET-Head: A lightweight auxiliary prediction head that achieves efficient multi-scale feature fusion through Ghost-based compression and adaptive gating with only 0.1M parameters and 0.02G FLOPs.

[0060] 2) MATR: A zero-parameter template optimization module that uses ternary metric discrimination to achieve occlusion recognition and performs quality-aware adaptive updates and occlusion recovery.

[0061] 3) GEM-Track: A high-performance plug-and-play tracking system that achieves the best accuracy-speed trade-off on multiple UAV benchmark datasets and has strong generalization capabilities for different ViT architectures.

[0062] Finally, experimental analysis was conducted on the embodiments of the present invention: This embodiment evaluates the GEM-Track tracker on five mainstream UAV tracking benchmark datasets: DTB70, UAV123, UAVDT, UAVTrack112, and UAVTrack112_L. DTB70 contains 70 dynamic aerial sequences, addressing challenges such as camera movement, cluttered backgrounds, and scale variations; UAV123 contains 123 low-altitude sequences, covering diverse scenes and targets; UAVDT consists of 50 vehicle sequences, focusing on evaluating target disappearance and occlusion; UAVTrack112 contains 112 real-world scene sequences, exceeding 100,000 frames, covering challenges such as fast movement and low resolution; UAVTrack112_L is its long-term subset, with each sequence exceeding 1000 frames, representing the largest-scale long-term aerial tracking benchmark for evaluating algorithm stability. Training was conducted on an AutoDL cloud instance (AMD EPYC 7642, 20 vCPUs, RTX 3090); evaluation was performed on a workstation equipped with an Intel i5-14600KF and an RTX 3090.

[0063] Further details on implementation: The GEM-Track tracker is built on SGLATrack-DeiT and uses DeiT-tiny as the backbone network. The template and search region sizes are set to 128×128 and 256×256 pixels, respectively. Key modifications include: adding a GET-Head as a lightweight parallel auxiliary branch in the prediction head; and integrating MATR during the inference phase to achieve adaptive template management.

[0064] The training data consisted of a combination of GOT-10k, LaSOT, COCO, and TrackingNet. The optimizer used was AdamW, with a batch size of 32 and weight decay of [value missing]. The predicted head learning rate is The backbone network learning rate is The learning rate was reduced by a factor of 10 at round 240. The training lasted for 300 rounds, with each round containing 60,000 sample pairs. A Hann window penalty was used during inference to suppress boundary effects.

[0065] Further comparison with the latest methods: This embodiment compares GEM-Track with 14 lightweight state-of-the-art trackers on five drone tracking benchmark datasets. Table 1 presents the AUC and Precision scores, and the average performance across the five datasets is expressed as "Avg.". GPU and CPU speeds are measured in frames per second (FPS).

[0066] Table 1. Comparison of data from five drone tracking benchmark datasets. ; As shown in Table 1, GEM-Track achieved the best average performance: AUC of 66.3% and Precision of 85.1%, exceeding the second-place SGLATrack-DeiT by 1.6% and 2.5%, respectively. This improvement stems from the synergistic effect of GET-Head and MATR: GET-Head enhances multi-scale localization capabilities through lightweight feature fusion, while MATR achieves adaptive template optimization through quality-aware updates.

[0067] The GGEM-Track tracker ranked first on four benchmark datasets. Notably, it achieved 66.5% AUC and 83.9% Precision on UAVTrack112_L, surpassing SGLATrack-DeiT by 2.5% and 4.7% respectively, validating the effectiveness of the dual-module design in suppressing long-term tracking accumulation error. On UAVDT, GEM-Track achieved 61.4% AUC and 84.9% Precision, demonstrating robustness against occlusion challenges.

[0068] Perform efficiency analysis: GEM-Track achieves an excellent accuracy-efficiency tradeoff: 164.87 frames per second on GPU and 35.35 frames per second on CPU. Compared to SGLATrack-DeiT, we trade a modest speed loss of approximately 19 frames per second for a significant accuracy improvement of 1.6% AUC and 2.5% Precision. This balance is achieved through a lightweight design: GET-Head only adds about 1.7% of parameters through GhostModule compression; MATR does not introduce any trainable parameters. Compared to ORTrack-DeiT, the GEM-Track tracker achieves a 90.7% speed improvement while maintaining higher accuracy. The CPU speed of 35.35 frames per second meets the real-time requirements of resource-constrained UAV platforms.

[0069] To further analyze the performance gains reported in Table 1, we present a detailed attribute-based analysis on the UAVTrack112 benchmark. The evaluation covers 13 challenging attributes: Low Resolution (LR), Aspect Ratio Change (ARC), Scale Variation (SV), Fast Motion (FM), Full Occlusion (FOC), Partial Occlusion (POC), Out-of-View (OV), Background Clutter (BC), Illumination Variation (IV), Viewpoint Change (VC), Camera Motion (CM), Similar Object (SOB), and Night / Low Illumination (N). Table 2 shows the AUC scores of the GEM-Track tracker and 14 state-of-the-art methods on these attributes.

[0070] Table 2 Attribute-based comparison data table on UAVTrack112 ; The results show that the GEM-Track tracker leads in 12 out of 13 attributes, demonstrating its comprehensive robustness. Notably, it achieves significant gains in scenarios involving appearance changes. For example, under illumination changes (IV), GEM-Track (70.22%) outperforms the second-place tracker by 4.5%; it also excels in scale changes (SV) and viewpoint changes (VC). This superior performance is primarily attributed to the synergy between GET-Head and MATR: GET-Head's enhanced feature representation provides a solid foundation for localization, while MATR's quality-aware mechanism ensures that the template is not corrupted by drastic appearance changes, thus maintaining long-term consistency.

[0071] Furthermore, the GEM-Track tracker demonstrates excellent robustness in dynamic and challenging scenarios. It ranked first in attribute tests for camera motion (CM, +2.63% compared to second place), fast motion (FM), and out-of-view (OV), highlighting the effectiveness of the dual-module design in maintaining tracking stability. MATR's adaptive optimization strategy prevents template drift during camera motion, while GET-Head's precise localization capability is crucial for re-detecting fast-moving or re-entering targets. These attribute-based experimental results validate that our proposed module effectively addresses key challenges in UAV tracking. In summary, this embodiment proposes two plug-and-play modules for optimizing a single-branch ViT tracker in UAV scenarios, focusing on both spatial and temporal dimensions. GET-Head achieves lightweight multi-scale feature fusion through Ghost-based compression and adaptive gating, requiring only 0.1M parameters and 0.02G FLOPs, thus improving robustness to scale variations and background interference. MATR implements quality-aware template updates under zero-parameter conditions through occlusion discrimination using ternary metrics and adaptive dynamic EMA fusion, helping to reduce accumulated errors in long-term tracking. Integrating these two modules into the single-branch ViT, we constructed the GEM-Track tracker, achieving a 1.6% improvement in AUC and a 2.5% improvement in precision on five UAV tracking benchmark datasets compared to state-of-the-art methods, while maintaining a real-time speed of 164.87 frames per second. The experiments demonstrate that these two modules are plug-and-play, continuously delivering performance gains across different ViT frameworks without requiring specific architectural adjustments.

[0072] Reference Figure 2 A drone tracking system based on a GEM-Track tracker includes: The first module 201 is used to acquire UAV video sequences and perform initialization and frame-by-frame tracking processing to obtain template tensors and search tensors. The second module 202 is used to construct the GEM-Track UAV tracker by introducing a multi-scale localization enhancement auxiliary prediction head and a metric perception optimization module based on the SGLATrack single-branch visual Transformer tracking architecture. The third module 203 is used to track template tensors and search tensors based on the GEM-Track UAV tracker, so as to realize real-time tracking of UAVs.

[0073] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0074] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A UAV tracking method based on a GEM-Track tracker, characterized in that, Includes the following steps: The drone video sequence is acquired and initialized and tracked frame by frame to obtain the template tensor and the search tensor. Based on the SGLATrack single-branch visual Transformer tracking architecture, a multi-scale localization enhancement auxiliary prediction head and a metric perception optimization module are introduced to construct the GEM-Track UAV tracker. Based on the GEM-Track drone tracker, the template tensor and search tensor are tracked to achieve real-time tracking of the drone.

2. The UAV tracking method based on a GEM-Track tracker according to claim 1, characterized in that, The step of acquiring the UAV video sequence and performing initialization and frame-by-frame tracking processing to obtain the template tensor and search tensor specifically includes: Acquire drone video sequences and extract the first frame image and its labeled target bounding box; Given a clipping factor, determine the clipping region using the target bounding box and calculate the side length of the clipping region; The first frame image is cropped according to the side length of the cropping region to obtain the template image; The template image is preprocessed using standardization to obtain the template tensor; Based on the second frame of the drone video sequence, frame-by-frame tracking is performed to obtain the original image of the current frame; Based on the target bounding box of the previous frame, a basic search cropping factor is set and adaptively adjusted. The original image of the current frame is cropped to obtain the search image. After standardization preprocessing, the search tensor is obtained.

3. The UAV tracking method based on a GEM-Track tracker according to claim 2, characterized in that, The GEM-Track UAV tracker specifically includes an initial template buffer, an online template buffer, a patch embedding layer, a ViT backbone network, a main prediction head, a multi-scale positioning enhancement auxiliary prediction head, and a metric perception optimization module.

4. The UAV tracking method based on a GEM-Track tracker according to claim 3, characterized in that, The step of tracking the template tensor and search tensor to achieve real-time tracking of the drone based on the GEM-Track drone tracker specifically includes: Input the template tensor and the search tensor into the GEM-Track drone tracker; An initial template buffer and an online template buffer based on the GEM-Track UAV tracker, wherein the initial template buffer and the online template buffer are used to store template tensors; Based on the patch embedding layer of the GEM-Track UAV tracker, the template tensor and search tensor of the online template buffer are divided into non-overlapping image blocks and mapped into feature vectors to obtain several template tags and several search tags. Learnable positional codes are added and concatenated to obtain a concatenated tag sequence. Based on the ViT backbone network of the GEM-Track UAV tracker, the search region features are separated and reshaped after ViT processing of the spliced ​​labeled sequence to obtain deep semantic features. The deep semantic features are first expanded in size using adaptive average pooling, and then restored in size using bilinear interpolation to obtain shallow features. Based on the main prediction head of the GEM-Track UAV tracker, multi-layer convolution processing is performed on deep semantic features to obtain center point score map, target size prediction map and sub-pixel offset prediction map. A multi-scale localization enhancement auxiliary prediction head based on the GEM-Track UAV tracker is used to perform auxiliary prediction on deep semantic features and shallow features to obtain an auxiliary confidence map. By combining the auxiliary confidence map, center point score map, target size prediction map and subpixel offset prediction map, score map fusion and bounding box decoding are performed to obtain the final fused score map, response map and target bounding box prediction result of the current frame; The metric perception optimization module based on the GEM-Track UAV tracker combines the target bounding box prediction results of the current frame to perform metric extraction and template fusion on the final fused score map and response map, thereby achieving real-time tracking of the UAV.

5. The UAV tracking method based on a GEM-Track tracker according to claim 4, characterized in that, The multi-scale localization enhancement auxiliary prediction head based on the GEM-Track UAV tracker performs auxiliary prediction on deep semantic features and shallow features to obtain an auxiliary confidence map. This step specifically includes: Deep semantic features and shallow features are input into the multi-scale localization enhancement auxiliary prediction head of the GEM-Track UAV tracker; The deep semantic features are processed by generating main features and auxiliary features respectively. The main features and auxiliary features are output and concatenated to obtain the aligned deep semantic features. Align the shallow features with channels and resolution to obtain aligned shallow features; Based on the global gating mechanism, the aligned deep semantic features and the aligned shallow features are combined into fusion candidate features. Dynamic weights are determined, and scalar gating weights and weighted deep weights are generated according to the dynamic weights. The aligned deep semantic features and the aligned shallow features are fused at multiple scales to obtain multi-scale gating fusion features. A channel recalibration mechanism is introduced to generate channel attention weights to recalibrate the multi-scale gated fusion features, thereby obtaining an auxiliary confidence map.

6. The UAV tracking method based on a GEM-Track tracker according to claim 5, characterized in that, The step of combining the auxiliary confidence map, center point score map, target size prediction map, and subpixel offset prediction map to perform score map fusion and bounding box decoding, and obtaining the final fused score map, response map, and target bounding box prediction result for the current frame, specifically includes: The auxiliary confidence map and the center point score map are fused with equal weights to obtain the final fused score map; A Hanning window is applied to the final fused score map to suppress edge responses, resulting in a response map. Based on the response map, the peak position of the response map is obtained. Combined with the target size prediction map and the sub-pixel offset prediction map, the normalized bounding box coordinates of the target are decoded. The occlusion count results from the metric-aware optimization module are used to adaptively adjust the basic search pruning factor. The target normalized bounding box coordinates are then combined to perform coordinate mapping, resulting in the target bounding box prediction for the current frame.

7. The UAV tracking method based on a GEM-Track tracker according to claim 6, characterized in that, The metric perception optimization module based on the GEM-Track UAV tracker, combining the target bounding box prediction results of the current frame, performs metric extraction and template fusion on the final fused score map and response map to achieve real-time tracking of the UAV. This step specifically includes: The final fused score map and response map are input into the metric perception optimization module of the GEM-Track drone tracker; Based on the quality metric extraction unit of the metric perception optimization module, the final fused score map and response map are metric extracted to obtain ternary metric information, which includes response entropy, kurtosis and confidence. An occlusion status discrimination mechanism and update conditions are set based on ternary metric information; If the update conditions are met, a new template image block is cropped based on the target bounding box prediction result of the current frame as a candidate template. The candidate template is then fused with the online template maintained in the previous time step using the exponential moving average algorithm to obtain the fused template image. Based on the occlusion counting results of the metric perception optimization module, the first frame image and the fused template image are introduced for anchor fusion to obtain a new fused template image, which is written back to the online template buffer. The target bounding box coordinates of each frame are output as the UAV tracking result.

8. A drone tracking system based on a GEM-Track tracker, characterized in that, Includes the following modules: The first module is used to acquire UAV video sequences and perform initialization and frame-by-frame tracking processing to obtain template tensors and search tensors. The second module is used to build the GEM-Track UAV tracker based on the SGLATrack single-branch visual Transformer tracking architecture by introducing a multi-scale localization enhancement auxiliary prediction head and a metric perception optimization module. The third module is used to track template tensors and search tensors based on the GEM-Track drone tracker, thereby enabling real-time tracking of drones.

Images