A multi-target tracking method and system for small targets of unmanned aerial vehicles

Abstract

The application discloses a kind of multi-target tracking method and system for unmanned aerial vehicle small target, belong to computer vision technical field.Method includes: based on video frame image after obtaining multistage feature map, multistage feature map is unified as initial fusion feature map;Based on initial fusion feature map and original image, double-dimension enhancement feature map is calculated;Based on the double-dimension enhancement feature map of two frames before and after, spatial enhancement feature map is calculated;After the double-dimension enhancement feature map of two frames before and after is spliced, in turn, feature alignment processing and memory pool interaction processing are carried out, current frame time enhancement feature map is obtained, and memory pool is updated;After spatial enhancement feature map and time enhancement feature map are fused, down-sampling and splicing are carried out, and the output result is further combined with ByteTrack tracking strategy, to realize multi-target tracking.The application method significantly improves accuracy and identity consistency, while maintaining efficient real-time inference performance.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a multi-target tracking method and system for small targets on drones, which can be applied to small target detection and tracking tasks in scenarios such as drone monitoring, intelligent transportation, and urban security. Background Technology

[0002] Multi-object tracking (MOT) is a fundamental task in computer vision, widely used in intelligent transportation, drone surveillance, and other scenarios. In recent years, deep learning-based detection algorithms like the YOLO series and tracking algorithms such as ByteTrack have achieved significant breakthroughs in medium to large target scenarios. However, tracking small targets (resolution below 32×32 pixels) in drone imagery still faces many challenges. Spatially, small targets have weak feature representations and low saliency, easily drowned out by background noise, making it difficult to form sufficiently discriminative features. Temporally, mainstream detectors often rely on single-frame inference, lacking modeling of inter-frame contextual information, leading to unstable detection results and easily interrupted target trajectories. In terms of confidence, complex backgrounds amplify the uncertainty of low-confidence features, causing false detections and reducing overall tracking stability.

[0003] Existing technologies also have significant limitations. For example, mainstream attention modules (such as SE-Net and CBAM) mainly focus on channel weighting or global feature modeling, making it difficult to accurately capture the local contrast between small targets and the background; multi-frame detection methods often rely on optical flow estimation or recurrent networks, which have high computational overhead and significantly degrade performance in low frame rates or fast motion scenes; in addition, most detectors do not fully consider the noise propagation problem in the feature enhancement process, resulting in a persistently high false detection rate. Summary of the Invention

[0004] This invention addresses the shortcomings of existing UAV small target multi-target tracking methods, such as insufficient spatial feature discriminativeness, poor temporal consistency, and high false detection rate. It provides a tracking method that integrates spatial augmentation and temporal modeling, improving the tracking accuracy and stability of small targets while ensuring real-time performance.

[0005] In a first aspect, the present invention proposes a multi-target tracking method for small targets on unmanned aerial vehicles, comprising:

[0006] S1. Obtain the video frame sequence to be processed captured by the drone. After preprocessing the original image of the current frame, further extract multi-level feature maps containing three different sizes. Unify the multi-level feature maps into the initial fusion feature map of the current frame through upsampling and stitching.

[0007] S2, based on the initial fused feature map of the current frame and the original image of the current frame, obtain the two-dimensional enhanced feature map of the current frame;

[0008] S3, based on the two-dimensional enhanced feature map of the previous frame and the two-dimensional enhanced feature map of the current frame, the motion difference of the current frame is calculated, and further based on the motion difference of the current frame and the detection confidence map of the current frame, the spatial enhanced feature map of the current frame is calculated; the detection confidence map of the current frame image is obtained by inputting the original image of the current frame into the detection head;

[0009] S4. After concatenating the two-dimensional enhanced feature map of the previous frame and the two-dimensional enhanced feature map of the current frame, feature alignment processing and memory pool interaction processing are performed in sequence to calculate the temporal enhanced feature map of the current frame and update the memory pool.

[0010] S5, after fusing the current frame spatial enhancement feature map and the current frame temporal enhancement feature map, downsampling and stitching are performed to obtain the current frame final enhancement feature map; the detection result is obtained based on the current frame final enhancement feature map; the detection result includes the detection box, category, detection confidence and corresponding feature vector of each target in the current frame image;

[0011] S6 combines detection results with the ByteTrack tracking strategy to achieve multi-target tracking through a two-stage matching mechanism.

[0012] Furthermore, in S1, multi-level feature maps are unified into the initial fused feature map of the current frame through upsampling and concatenation, specifically as follows:

[0013] The multi-level feature maps are labeled as dark3, dark4, and dark5 according to their size from low to high. Dark5 is upsampled once and then concatenated with dark4. After being fused using CSPDarknet, it is upsampled once and then concatenated with dark3 to obtain the initial fused feature map.

[0014] Furthermore, the specific steps of S2 are as follows:

[0015] S201, the initial fused feature map is passed sequentially through a depthwise separable convolutional unit and a SimAM channel attention unit to obtain a channel attention map;

[0016] S202, add the feature attention map to the corresponding original image to obtain the output feature;

[0017] S203, the output features are scaled and local contrast enhanced to calculate the channel enhanced feature map and spatial attention mask respectively;

[0018] S204: Multiply the channel enhancement feature map pixel by pixel with the spatial attention mask, scale it by introducing a scaling factor, and then add it to the channel enhancement feature map to obtain the current frame's two-dimensional enhancement feature map.

[0019] Furthermore, the depth-separable convolutional unit is composed of two DWConv units connected in series; the local contrast enhancement unit is composed of one DWconv unit.

[0020] Furthermore, in S3, the specific calculation formula for the spatial enhancement feature map of the current frame is as follows:

[0021] ;

[0022] in, Let (x, y) be the value of the pixel on the detection confidence map. To detect the mean of the confidence plot, For modulation intensity, Due to differences in movement, This enhances spatial features.

[0023] Furthermore, the specific formula for calculating the motion difference is as follows:

[0024] ;

[0025] in, For the two-dimensional enhancement features of the previous frame, It is a depthwise separable convolution function. The motion sensitivity threshold, For the current frame, two-dimensional enhancement features, This is the sigmoid function.

[0026] Furthermore, in S4, the memory pool is used to store feature embedding vectors and detection confidence maps of several historical frames; the specific calculation steps for the temporal augmentation feature map are as follows:

[0027] S401, input the two-dimensional enhanced feature map of the previous frame and the two-dimensional enhanced feature map of the current frame into a single gated convolutional layer to obtain the feature map of the current frame after feature alignment;

[0028] S402, the QKV attention mechanism is used to extract the query matrix from the feature map of the current frame after feature alignment, and the feature embedding vectors of all historical frames in the memory pool are used to calculate the feature map of the current frame after attention interaction.

[0029] S403: Multiply the current frame feature map after attention interaction and the current frame dual-dimensional enhanced feature map pixel by pixel, and then further superimpose the current frame dual-dimensional enhanced feature map to calculate the current frame temporal enhanced feature map.

[0030] Furthermore, in S4, the update mechanism of the memory pool is as follows: if the mean value of the detection confidence map of the current frame is greater than the mean value of the detection confidence maps of all existing historical frames in the memory pool, then an update is triggered, and the feature embedding vector and detection confidence of the current frame are written into the memory pool in a sliding window manner; otherwise, the calculation ends.

[0031] Furthermore, the two-stage matching mechanism is specifically as follows:

[0032] For the starting frame of the video frame sequence, based on the set threshold and the detection confidence of each target, high-confidence detection boxes are selected and initialized as the initial trajectory corresponding to the target;

[0033] For non-starting frames in a video frame sequence, high-confidence and low-confidence detection boxes are obtained by filtering based on a set threshold and the detection confidence of each target. First, the feature vectors corresponding to the high-confidence detection boxes are used to calculate the optimal trajectory with the initial trajectory, and the detection boxes are associated with and supplemented into the corresponding optimal initial trajectories. Second, the feature vectors of the unassociated initial trajectories are matched with the feature vectors of the low-confidence detection boxes for similarity. If the similarity is higher than the threshold, the detection box is associated with the corresponding initial trajectory. If the similarity is lower than the threshold, the calculation ends.

[0034] Secondly, this invention proposes a multi-target tracking system for small targets of UAVs, which is used to implement the above-mentioned multi-target tracking method for small targets of UAVs.

[0035] The beneficial effects of this invention are:

[0036] (1) In terms of spatial feature enhancement, the contrast perception attention mechanism is used to achieve dual-dimensional optimization of channel semantics and local spatial contrast, and the small target enhancement path is combined to effectively suppress background noise, thereby significantly improving the discriminativeness of small target features and reducing the false negative rate.

[0037] (2) In terms of improving temporal consistency, the lightweight inter-frame alignment and historical feature memory of the temporal enhancement module can compensate for feature degradation caused by occlusion and motion blur, reduce the number of identity switching, and improve the continuity of the target trajectory.

[0038] (3) In terms of the balance between accuracy and real-time performance, this method significantly improves the accuracy and identity consistency of UAV small target multi-target tracking, while maintaining efficient real-time inference performance, and can be well adapted to resource-constrained platforms such as UAVs. Attached Figure Description

[0039] Figure 1 This is a flowchart of the spatiotemporal augmentation network (TASONet) for multi-target tracking method for small targets of UAVs according to the present invention. Detailed Implementation

[0040] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0041] This invention proposes a multi-target tracking method for small targets on unmanned aerial vehicles (UAVs), implemented based on a temporal-spatial augmentation network (TASONet), including steps such as feature extraction, spatial augmentation, temporal augmentation, feature aggregation, and tracking output. The specific implementation steps are as follows:

[0042] Step 1: Construction and Training of Spatiotemporal Augmentation Network

[0043] The spatiotemporal augmentation network (TASONet) constructed in this invention is an improved YOLOPAFPN architecture based on CSPDarknet as the backbone network. It mainly includes modules such as CSPDarknet for feature extraction, small target augmentation module, temporal augmentation module, path aggregation network, detection head and identity recognition branch.

[0044] Based on the UAVDT and VisDrone-MOT datasets, and using the intersection-over-union ratio (IoU) as the loss function, this invention further uses gradient descent to train each module in the spatiotemporal augmentation network, ultimately obtaining the trained spatiotemporal augmentation network.

[0045] Step 2: Multi-level feature extraction and initial fusion

[0046] After acquiring the video captured by the drone, each frame of the video is preprocessed, including scaling and normalization, and then input into the CSPDarknet backbone network. CSPDarknet extracts multi-level feature maps containing low-level structural information and high-level semantic information (feature map sizes from low to high are dark3, dark4, and dark5, corresponding to different levels of features).

[0047] Multi-level feature maps are fused from top to bottom using a Feature Pyramid Network (FPN). The features of dark5 are upsampled once and concatenated with the features of dark4, then fused through CSPDarknet. After that, upsampling continues and concatenation with the features of dark3 to obtain the initial fused feature map. This step is to allow deep semantic features to flow to the middle and shallow layers, providing basic feature input for subsequent small target enhancement.

[0048] Step 3: Spatial Feature Enhancement and Noise Suppression

[0049] The initial fused feature map obtained in step one is input into the small object enhancement module (SOE). Adaptive noise suppression is achieved by calculating the motion differences between consecutive frames. The small object enhancement module includes units such as depthwise convolution thinning, channel attention enhancement, local contrast enhancement, and small object path enhancement.

[0050] The specific steps include:

[0051] (3.1) Depth-separable convolutional refinement: Two DWConv (3×3) are concatenated to form a depth-separable convolution, which refines the local space of the initial fused feature map of the input to extract fine-grained spatial details, reduce the amount of computation while preserving the integrity of the features, and provide more stable local structural information for the subsequent attention module;

[0052] (3.2) Contrastive Perception Attention Optimization: Based on (2.1), SimAM channel attention enhancement is first applied to the refined features. The channel attention map is obtained by calculating the energy function of each pixel within each channel, thereby filtering channels with high semantic concentration and suppressing noise-dominant channels. The specific formula for calculating the energy function is as follows:

[0053]

[0054] in, The mean of the features within the channel. The characteristic variance within the channel, is the regularization constant, E(z) is the energy of each pixel, and z is the pixel value of the pixel in the original image.

[0055] Secondly, the channel attention map is added to the corresponding original image to obtain the output feature. After the scaling parameter is adjusted, the output feature becomes the channel enhancement feature, which further improves the response of the salient region of the small target.

[0056] The formula for calculating channel enhancement features is:

[0057]

[0058] in, For scaling parameters, Here, F represents the channel weights, and F represents the output features. To enhance channel features, This is for element-wise multiplication.

[0059] Secondly, local neighborhood mean and variance are calculated through LCDAttention local contrast enhancement to generate local contrast encoding and adaptive spatial attention mask, enhancing the brightness and texture differences between small targets and the background, and outputting two-dimensional enhanced features. Specifically, LCDAttention local contrast enhancement employs an LCDAttention module based on local statistics, which consists of a single DWconv depthwise convolution, constructing a feature for each pixel in the output. The local mean and variance of the neighborhood (in this specific embodiment, the neighborhood size is set to 3×3) are used to obtain the Local Contrast Encoding (LCE). The specific calculation formula for the Local Contrast Encoding is as follows:

[0060]

[0061] in, for The local mean of the neighborhood. for Local variance of the neighborhood It is the stability constant. The confidence level that a target exists in each local region. For pixels The feature vector on the output feature F.

[0062] By inputting the LCE into a convolutional sigmoid attention structure, an adaptive spatial attention mask is obtained. This mask enhances high-contrast regions (typically small targets) while suppressing background regions, thus achieving two-dimensional optimization of the channel space. (Adaptive Spatial Attention Mask) The calculation formula is as follows:

[0063]

[0064] in, for Convolution kernel, This is the fusion ratio coefficient. This is the sigmoid function.

[0065] Finally, the enhanced features are calculated as follows:

[0066]

[0067] in, This is a learnable scaling factor.

[0068] (3.3) Small target enhancement path noise reduction: This step is used to enhance the modeling ability of the shallowest layer features for small targets. By calculating the motion difference between the features of the current frame and the previous frame, and combining the detection confidence C output by the detection head based on the current frame image, adaptive noise suppression is achieved. The response of low confidence regions is weakened, and the features of high confidence small target regions are enhanced, ultimately obtaining spatial enhancement features.

[0069] The formula for calculating motion differences is:

[0070]

[0071] in, This is an enhancement feature of the previous frame. For depthwise separable convolution, The motion sensitivity threshold, Feature map with added motion differences.

[0072] Based on this, the spatial enhancement features are calculated according to the detection confidence map. The specific calculation formula is as follows:

[0073]

[0074] in, Let (x, y) be the value of the pixel on the detection confidence map. To detect the global mean of the confidence plot, For modulation intensity, For spatial enhancement feature maps.

[0075] Step 4: Temporal Feature Alignment and Memory Modeling

[0076] The continuous frame features output by the backbone network The input is fed into the Temporal Augmentation Module (TEM), which consists of a gated convolutional layer and a memory pool of a certain size. It includes two sub-units: a temporal feature alignment unit and a target memory unit, used to capture inter-frame consistency and reduce short-term feature breaks caused by jitter and occlusion. Specific steps include:

[0077] (4.1) Temporal feature alignment: The features of the current frame and the historical frames are concatenated by the gated convolution of the temporal feature alignment unit, and the alignment features are generated by 3×3 convolution; at the same time, the gate weight (sigmoid) is generated from the current frame features to control the fusion ratio of the alignment features, so as to achieve lightweight inter-frame alignment. When there is no historical frame input (such as the first frame), the current frame features are output.

[0078] This step dynamically adjusts the fusion weights of the current frame and historical frame features based on the inter-frame motion amplitude, achieving lightweight inter-frame feature alignment without optical flow estimation.

[0079] The specific formula for feature alignment is as follows:

[0080]

[0081] in, For gated convolution kernels, To align the convolution kernels, For the sigmoid function, For element-wise multiplication, Enhanced features after feature alignment; Used to generate gating weights.

[0082] (4.2) Target memory interaction: A sliding memory pool of a certain size is set inside the target memory unit (in this specific embodiment, the memory pool capacity is 20 frames). This memory pool is used to store the feature embedding vectors and detection confidence corresponding to high-confidence frames in the historical frames; through a self-attention mechanism, the features of the current frame are... QKV attention is employed to interact with features in the memory pool, compensating for feature degradation caused by occlusion and motion blur, and resulting in temporally enhanced features associated with historical information. The attention interaction formula is as follows:

[0083]

[0084] Where Q is the feature of the current frame. The query matrix, K and V are the key-value matrices of the memory pool confidence, and N is the number of feature channels. Enhance the features of the current frame after attention interaction.

[0085] After forward propagation, the memory pool is further updated based on the confidence level. The update condition for the memory pool is as follows: ,in This represents the mean of the detection confidence map for the current frame. The mean of the detection confidence map of all historical frames in the memory pool is used for this update mechanism to ensure that only high-confidence samples are written to the memory pool, thereby avoiding background noise contaminating long-term memory.

[0086] (4.3) Calculation of temporal augmentation features. Based on the above calculations, the final temporal augmentation features are obtained. The specific calculation formula is as follows:

[0087]

[0088] Step 5: Feature Aggregation and Tracking Output

[0089] spatial enhancement features With time-enhanced features The input is fed into the Path Aggregation Network (PAN) for bottom-up secondary fusion. Two downsampling operations are performed sequentially, downsampling the feature map size from dark3 to dark4 and then to dark5. Each downsampling result is concatenated with the corresponding mid- and high-level features of the FPN to generate an enhanced feature map that is semantically and spatially consistent across scales.

[0090] The enhanced feature map is fed into both the detection head and the identity recognition branch. The detection head predicts the target detection box, category, and confidence level. The identity recognition branch extracts a highly discriminative feature vector for each target and combines it with the historical frame features stored in the memory pool to perform similarity matching for each target.

[0091] Combining the ByteTrack tracking strategy, multi-target tracking is achieved frame-by-frame through a two-stage matching mechanism. The specific matching steps are as follows: Detection boxes with confidence levels above a threshold are designated as high-confidence detection boxes. For the initial frame of the sequence, these high-confidence detection boxes are initialized as the initial trajectory of the corresponding target. For the remaining video frames in the sequence, after the high-confidence detection boxes of the current frame are used to associate with the initial trajectories, the similarity matching results between the unmatched initial trajectories and the low-confidence detection boxes are calculated, and trajectory completion and matching are performed.

[0092] Repeat the above steps until all video frames in the sequence have been traversed, and finally output a complete tracking trajectory that is spatiotemporally continuous and consistent in identity, thus realizing multi-target tracking of small targets by UAVs.

[0093] To verify the effectiveness and superiority of the method of the present invention, a comparative experiment was further conducted based on the two major UAV small target tracking benchmark datasets, UAVDT and VisDrone-MOT.

[0094] The experimental hardware environment consisted of an Intel Core i7-6850K CPU and an NVIDIA RTX 1080ti GPU, while the software environment was PyTorch 1.10. The core evaluation metrics were MOTA (Multi-Object Tracking Accuracy), IDF1 (Identity Consistency Score), FP+FN (Total False Positives and False Negatives), IDs (Identity Switching Count), and FPS (FPS) (Real-Time Inference Speed). The results were compared with mainstream algorithms such as ByteTrack and BoTSORT. The experimental results are shown in Tables 1 and 2. It can be seen that the method of this invention achieves a MOTA of 75.97% and an IDF1 of 88.51% on the UAVDT dataset; and a MOTA of 73.52% and an IDF1 of 88.83% on the VisDrone-MOT dataset, while maintaining a real-time inference speed of approximately 12 FPS. This demonstrates efficient and real-time inference performance while maintaining accuracy.

[0095] Table 1. Performance of each method on the UAVDT dataset

[0096]

[0097] Table 2. Performance of each method on the VisDrone-MOT dataset

[0098]

[0099] Based on the same inventive concept, this invention also provides a multi-target tracking system for small targets on unmanned aerial vehicles (UAVs), used to implement the aforementioned multi-target tracking method for small targets on UAVs. The system includes:

[0100] The upsampling module is used to acquire the video frame sequence to be processed captured by the drone. After preprocessing the original image of the current frame, it further extracts multi-level feature maps containing three different sizes. The multi-level feature maps are unified into the initial fusion feature map of the current frame through upsampling and stitching.

[0101] The small target enhancement module is used to obtain a two-dimensional enhanced feature map of the current frame based on the initial fused feature map of the current frame and the original image of the current frame;

[0102] The spatial enhancement feature module is used to calculate the motion difference of the current frame based on the two-dimensional enhancement feature map of the previous frame and the two-dimensional enhancement feature map of the current frame, and further calculate the spatial enhancement feature map of the current frame based on the motion difference of the current frame and the detection confidence map of the current frame; the detection confidence map of the current frame image is obtained by inputting the original image of the current frame into the detection head;

[0103] The temporal enhancement module is used to concatenate the two-dimensional enhanced feature map of the previous frame and the two-dimensional enhanced feature map of the current frame, and then perform feature alignment processing and memory pool interaction processing in sequence to calculate the temporal enhanced feature map of the current frame and update the memory pool.

[0104] The final enhanced feature calculation module is used to fuse the current frame spatial enhanced feature map and the current frame temporal enhanced feature map, then downsample and stitch them together to obtain the current frame final enhanced feature map; the detection result is obtained based on the current frame final enhanced feature map; the detection result includes the detection box, category, detection confidence and corresponding feature vector of each target in the current frame image;

[0105] The multi-target tracking module combines detection results with the ByteTrack tracking strategy to achieve multi-target tracking through a two-stage matching mechanism.

[0106] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0107] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A multi-target tracking method for small targets on unmanned aerial vehicles (UAVs), characterized in that, include: S1. Obtain the video frame sequence to be processed captured by the drone. After preprocessing the original image of the current frame, further extract multi-level feature maps containing three different sizes. Unify the multi-level feature maps into the initial fusion feature map of the current frame through upsampling and stitching. S2, based on the initial fused feature map of the current frame and the original image of the current frame, obtain the two-dimensional enhanced feature map of the current frame; S3, based on the two-dimensional enhanced feature map of the previous frame and the two-dimensional enhanced feature map of the current frame, the motion difference of the current frame is calculated, and further based on the motion difference of the current frame and the detection confidence map of the current frame, the spatial enhanced feature map of the current frame is calculated; the detection confidence map of the current frame image is obtained by inputting the original image of the current frame into the detection head; S4. After concatenating the two-dimensional enhanced feature map of the previous frame and the two-dimensional enhanced feature map of the current frame, feature alignment processing and memory pool interaction processing are performed in sequence to calculate the temporal enhanced feature map of the current frame and update the memory pool. S5, after fusing the current frame spatial enhancement feature map and the current frame temporal enhancement feature map, downsampling and stitching are performed to obtain the current frame final enhancement feature map; the detection result is obtained based on the current frame final enhancement feature map; the detection result includes the detection box, category, detection confidence and corresponding feature vector of each target in the current frame image; S6 combines detection results with the ByteTrack tracking strategy to achieve multi-target tracking through a two-stage matching mechanism.

2. The multi-target tracking method for small targets on unmanned aerial vehicles according to claim 1, characterized in that, In S1, multi-level feature maps are unified into the initial fused feature map of the current frame through upsampling and concatenation, specifically: The multi-level feature maps are labeled as dark3, dark4, and dark5 according to their size from low to high. Dark5 is upsampled once and then concatenated with dark4. After being fused using CSPDarknet, it is upsampled once and then concatenated with dark3 to obtain the initial fused feature map.

3. The multi-target tracking method for small targets on unmanned aerial vehicles according to claim 1, characterized in that, The specific steps of S2 are as follows: S201, the initial fused feature map is passed sequentially through a depthwise separable convolutional unit and a SimAM channel attention unit to obtain a channel attention map; S202, add the feature attention map to the corresponding original image to obtain the output feature; S203, the output features are scaled and local contrast enhanced to calculate the channel enhanced feature map and spatial attention mask respectively; S204: Multiply the channel enhancement feature map pixel by pixel with the spatial attention mask, scale it by introducing a scaling factor, and then add it to the channel enhancement feature map to obtain the current frame's two-dimensional enhancement feature map.

4. The multi-target tracking method for small targets on unmanned aerial vehicles according to claim 3, characterized in that, The depthwise separable convolutional unit is composed of two DWConv units connected in series; the local contrast enhancement unit is composed of one DWconv unit.

5. A multi-target tracking method for small targets on unmanned aerial vehicles according to claim 1, characterized in that, In S3, the specific calculation formula for the spatial enhancement feature map of the current frame is as follows: ； in, Let (x, y) be the value of the pixel on the detection confidence map. To detect the mean of the confidence plot, For modulation intensity, Due to differences in movement, For spatial enhancement features, This is for element-wise multiplication.

6. A multi-target tracking method for small targets on unmanned aerial vehicles according to claim 5, characterized in that, The specific formula for calculating the motion difference is as follows: ； in, For the two-dimensional enhancement features of the previous frame, It is a depthwise separable convolution function. The motion sensitivity threshold, For the current frame, two-dimensional enhancement features, This is the sigmoid function.

7. A multi-target tracking method for small targets on unmanned aerial vehicles according to claim 1, characterized in that, In S4, the memory pool is used to store the feature embedding vectors and detection confidence maps of several historical frames; the specific calculation steps for the temporal augmentation feature map are as follows: S401, input the two-dimensional enhanced feature map of the previous frame and the two-dimensional enhanced feature map of the current frame into a single gated convolutional layer to obtain the feature map of the current frame after feature alignment; S402, the QKV attention mechanism is used to extract the query matrix from the feature map of the current frame after feature alignment, and the feature embedding vectors of all historical frames in the memory pool are used to calculate the feature map of the current frame after attention interaction. S403: Multiply the current frame feature map after attention interaction and the current frame dual-dimensional enhanced feature map pixel by pixel, and then further superimpose the current frame dual-dimensional enhanced feature map to calculate the current frame temporal enhanced feature map.

8. A multi-target tracking method for small targets on unmanned aerial vehicles according to claim 1, characterized in that, In S4, the update mechanism of the memory pool is as follows: if the mean value of the detection confidence map of the current frame is greater than the mean value of the detection confidence maps of all existing historical frames in the memory pool, then an update is triggered, and the feature embedding vector and detection confidence of the current frame are written into the memory pool in a sliding window manner; otherwise, the calculation ends.

9. A multi-target tracking method for small targets on unmanned aerial vehicles according to claim 1, characterized in that, The two-stage matching mechanism is as follows: For the starting frame of the video frame sequence, based on the set threshold and the detection confidence of each target, high-confidence detection boxes are selected and initialized as the initial trajectory corresponding to the target; For non-starting frames in a video frame sequence, high-confidence and low-confidence detection boxes are obtained by filtering based on a set threshold and the detection confidence of each target. First, the feature vectors corresponding to the high-confidence detection boxes are used to calculate the optimal trajectory with the initial trajectory, and the detection boxes are associated with and supplemented into the corresponding optimal initial trajectories. Second, the feature vectors of the unassociated initial trajectories are matched with the feature vectors of the low-confidence detection boxes for similarity. If the similarity is higher than the threshold, the detection box is associated with the corresponding initial trajectory. If the similarity is lower than the threshold, the calculation ends.

10. A multi-target tracking system for small targets on unmanned aerial vehicles (UAVs), used to implement the multi-target tracking method for small targets on UAVs as described in claim 1, characterized in that, The system includes: The upsampling module is used to acquire the video frame sequence to be processed captured by the drone. After preprocessing the original image of the current frame, it further extracts multi-level feature maps containing three different sizes. The multi-level feature maps are unified into the initial fusion feature map of the current frame through upsampling and stitching. The small target enhancement module is used to obtain a two-dimensional enhanced feature map of the current frame based on the initial fused feature map of the current frame and the original image of the current frame; The spatial enhancement feature module is used to calculate the motion difference of the current frame based on the two-dimensional enhancement feature map of the previous frame and the two-dimensional enhancement feature map of the current frame, and further calculate the spatial enhancement feature map of the current frame based on the motion difference of the current frame and the detection confidence map of the current frame; the detection confidence map of the current frame image is obtained by inputting the original image of the current frame into the detection head; The temporal enhancement module is used to concatenate the two-dimensional enhanced feature map of the previous frame and the two-dimensional enhanced feature map of the current frame, and then perform feature alignment processing and memory pool interaction processing in sequence to calculate the temporal enhanced feature map of the current frame and update the memory pool. The final enhanced feature calculation module is used to fuse the current frame spatial enhanced feature map and the current frame temporal enhanced feature map, then downsample and stitch them together to obtain the current frame final enhanced feature map; the detection result is obtained based on the current frame final enhanced feature map; the detection result includes the detection box, category, detection confidence and corresponding feature vector of each target in the current frame image; The multi-target tracking module combines detection results with the ByteTrack tracking strategy to achieve multi-target tracking through a two-stage matching mechanism.

Images