A gradient layered multi-branch air-to-air unmanned aerial vehicle target detection method

Abstract

The application belongs to the technical field of target detection, and discloses a gradient layered multi-branch air-to-air unmanned aerial vehicle target detection method, which specifically comprises the following steps: S1, preparing a data set; S2, designing a multi-branch down-sampling module. The application designs a gradient layered multi-branch air-to-air unmanned aerial vehicle target detection method, designs a structure network suitable for air-to-air unmanned aerial vehicle target detection, makes the network suitable for processing complex background interference in the air-to-air scene, realizes multi-scale adaptive focusing, designs a multi-branch down-sampling module, performs gradient layered design according to the resolution reduction stage, realizes differentiated extraction of multi-scale features, designs a space-time-scale perception ST-C2f feature extraction module, integrates a space-time attention mechanism, enhances feature extraction of moving targets, optimizes a suitable loss function, achieves multi-scale balance, gradient stability, and reduces missed detection and other advantages; experimental data proves that the application shows good performance in the target detection task.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, specifically to a gradient-layered multi-branch air-to-air unmanned aerial vehicle (UAV) target detection method. Background Technology

[0002] With the continuous evolution of information and intelligent technologies, drones, with their low cost, small size, and ease of operation, have largely compensated for the shortcomings of traditional image acquisition methods. Compared with fixed cameras, the mobility of drones provides greater flexibility to air-to-air perspectives, enabling them to capture scenes that are difficult to cover with ground-based photography. This has driven the continuous deepening of research in the field of drone target detection and tracking, with target detection algorithms being one of the core research directions in the current field of computer vision. However, existing detection technologies struggle to cope with complex and ever-changing environments. Against this backdrop, air-to-air drone photography shows promising application prospects. However, due to the high altitude of drones and the greater complexity of the subjects being photographed, targets in the images tend to be smaller and move faster, posing a greater challenge to the target detection capabilities of existing algorithms. Therefore, this invention proposes a gradient-layered multi-branch air-to-air drone target detection method, constructing a detection network suitable for air-to-air drone targets.

[0003] Chinese patent publication number "CN120198822A" entitled "A Small Target Detection Algorithm for UAVs Based on Improved YOLOv8" describes an algorithm that, through a nonlocal mechanism and the introduction of a dynamic CBAM module, enables the model to automatically adjust its feature focus area based on the specific features of each image. During feature extraction, by modeling the global pixel correlation in the feature map, compared to modules relying on local pixel correlation, the features of small targets can be recognized from a global perspective, helping the model more effectively focus on small, easily confused target areas. However, its network lacks design considerations for fast-moving targets, making it unsuitable for air-to-air UAV target detection scenarios and exhibiting poor generalization ability. Furthermore, relying solely on the CBAM module is insufficient for adapting to dynamic scenes and complex backgrounds. Therefore, designing a target detection method with better adaptability to dynamic scenes and complex backgrounds in air-to-air UAV environments, achieving higher accuracy, and being more suitable for target detection in air-to-air UAV applications is a direction urgently requiring research in this invention.

[0004] Therefore, we propose a gradient-layered multi-branch air-to-air UAV target detection method to solve the above problems. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a gradient-layered multi-branch air-to-air UAV target detection method, which solves the problems of low detection accuracy, inability to adapt to complex target backgrounds, and lack of perception of dynamic scene information in the air-to-air UAV target detection process.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0009] A gradient-layered multi-branch air-to-air UAV target detection method, which specifically includes the following steps:

[0010] S1, Prepare the dataset: Prepare an image dataset of air-to-air UAV targets, including images and extracted optical flow information;

[0011] S2, Design of a multi-branch downsampling module: This module includes a three-branch downsampling module and a two-branch downsampling module, which contain standard convolution, deformable convolution, depthwise separable convolution, BN layer, max pooling and activation function;

[0012] S3, Design of the spatiotemporal scale-aware ST-C2f feature extraction module: This module includes a spatiotemporal attention mechanism, standard convolution, deformable convolution, and the convolutional layers contain normalization and activation functions, which are finally connected and combined through the designed residual structure;

[0013] S4, Design the loss function: Design a composite loss function, including classification loss, bounding box regression loss, target confidence loss and motion perception loss;

[0014] S5, Construct and train the object detection network model: Construct a new object detection network by adding a multi-branch downsampling module and a spatiotemporal scale-aware ST-C2f feature extraction module. The object detection network model includes a backbone network, a neck network and four detection heads.

[0015] S6, Select Evaluation Metrics: Select the best evaluation metrics to measure the accuracy of the gradient hierarchical multi-branch air-to-air UAV target detection method and evaluate the target detection results completed in S5.

[0016] Furthermore, in S2, the multi-branch downsampling module adjusts the number and type of branches according to the characteristics of different resolution layers. First, a three-branch structure is used in the high-resolution layer, and a two-branch structure is used in the low-resolution layer. This can achieve comprehensive capture of multi-scale features and efficient allocation of computing resources, improve accuracy, and is suitable for air-to-air UAV detection application scenarios.

[0017] The three-branch structure includes standard convolution max pooling and depthwise separable convolution, which are used to capture local features while preserving texture edge information and preserve spatial structure, respectively.

[0018] The dual-branch structure includes standard convolution and depthwise separable convolution, which are used to capture semantic features and perform lightweight downsampling, respectively, balancing accuracy and efficiency.

[0019] Furthermore, in S3, the connection method of each structure of the spatiotemporal scale-aware ST-C2f feature extraction module is as follows: the input data is first passed through a spatiotemporal attention module and then enters the first convolutional layer, and then input into the bottleneck module 3. Then, it is residually connected with the bottleneck module 1, bottleneck module 2 and bottleneck module 3 respectively. The outputs from bottleneck module 3 to bottleneck module 4 and from bottleneck module 1 to bottleneck module 2 are concatenated into the second convolutional layer and finally output. Bottleneck modules 1 and 2 are composed of a convolutional layer, a deformable convolutional layer and a convolutional layer respectively. Bottleneck modules 3 and 4 are composed of variable convolutional layers respectively. This structure improves the feature extraction capability, improves the accuracy of the target detection network, and significantly enhances the adaptability to aerial scenes. In addition, passing the spatiotemporal scale-aware ST-C2f feature extraction module through a C2f module and SPPF module in the high and medium resolution layers is the optimal strategy that balances performance and efficiency, which can maximize the detection accuracy of dynamic targets and adapt to the detection needs of multiple scales.

[0020] The spatiotemporal attention module compresses high-dimensional optical flow information and aligns it with the feature map by fusing spatial and temporal dimension information. Then, it splices the original feature map with the preprocessed optical flow field in the channel dimension. It is suitable for dynamic target detection tasks in air-to-air scenarios and has efficient spatiotemporal information fusion and dynamic attention to moving regions.

[0021] Furthermore, in S4, the loss function during training includes four parts: classification loss, bounding box regression loss, target confidence loss, and motion perception loss.

[0022] The classification loss function adopts an improved version of Focal Loss, which alleviates class imbalance through dynamic weights. The bounding box regression loss function adopts a combination of GIoU and DIoU loss functions. The target confidence loss adopts a weighted binary cross-entropy. The motion perception loss includes two parts: optical flow consistency loss function and temporal feature difference loss function. This loss function has the advantages of multi-scale adaptation, dynamic scene optimization, good module collaboration and improved robustness.

[0023] Furthermore, in S5, the new target detection network backbone network goes through three spatiotemporal scale-aware ST-C2f feature extraction modules, one C2f module, one SPPF module, and a multi-branch downsampling module. Then, through the neck network, it undergoes top-down and bottom-up feature fusion and feature extraction to obtain a multi-scale information feature map. Finally, four detection heads are added.

[0024] The four detection heads are used to detect large targets, medium targets, small targets, and micro targets, respectively.

[0025] (III) Beneficial Effects

[0026] Compared with existing technologies, this invention provides a gradient-layered multi-branch air-to-air UAV target detection method, which has the following beneficial effects:

[0027] This invention provides a gradient-layered multi-branch air-to-air UAV target detection method. It designs a new target detection network with gradient-layered multi-branched multi-scale fusion. The target detection network includes a backbone network, a neck network, and four detection heads, which is more suitable for air-to-air UAV target detection. This addresses the problems of low detection accuracy, inability to adapt to complex target backgrounds, and lack of perception of dynamic scene information in the air-to-air UAV target detection process.

[0028] This invention designs a spatiotemporal scale-aware ST-C2f feature extraction module in the target detection network structure. It integrates spatial features and optical flow information through a spatiotemporal attention mechanism and adopts a dynamic branching strategy to adapt to multi-scale targets. Under a lightweight design, it breaks through the local receptive field limitation of traditional convolution, while enhancing the consistency of temporal features and improving adaptability to dynamic scenes. It solves the problems of traditional air-to-air target detection not being suitable for complex target backgrounds and lacking perception of dynamic scene information.

[0029] This invention designs a gradient hierarchical multi-branch downsampling module. By dynamically adjusting the number of branches and fusing complementary features of standard convolution, pooling, and depthwise separable convolution, it retains richer multi-scale information during dimensionality reduction, overcomes the feature loss problem of traditional single downsampling operations, enhances gradient propagation stability, and improves the feature representation capability for targets of different scales. As a result, it improves the detection accuracy in the air-to-air UAV target detection process, especially for target detection in complex scenarios where the gain is more significant.

[0030] This invention designs a novel composite loss function that uses classification loss, bounding box regression loss, target confidence loss, and motion perception loss to jointly constrain network training. The classification loss optimizes the prediction accuracy of target categories, adapting to the multi-scale characteristics of four detector heads; the bounding box regression loss, combined with optical flow compensation, improves the accuracy of localizing moving targets; the target confidence loss strengthens foreground-background differentiation and adds temporal consistency constraints; and the motion perception loss ensures the temporal stability of dynamic target features. Therefore, the composite loss function involved in this invention can effectively improve the detection performance of targets at multiple scales and in dynamic scenes.

[0031] The gradient hierarchical multi-branch air-to-air UAV target detection method designed in this invention has a certain scene generalization ability. This algorithm performs well in different scenarios and has strong scene adaptability. It has good adaptability and generalization ability for target detection in different scenarios. Attached Figure Description

[0032] Figure 1 This is a flowchart of a gradient-layered multi-branch air-to-air unmanned aerial vehicle target detection method according to the present invention;

[0033] Figure 2 This is a schematic diagram of the target detection network structure constructed in this invention;

[0034] Figure 3 This is a schematic diagram of the specific structure of the backbone network of the present invention;

[0035] Figure 4 This is a schematic diagram of the multi-branch downsampling module of the present invention;

[0036] Figure 5 This is a schematic diagram of the spatiotemporal scale-aware ST-C2f feature extraction module designed in this invention;

[0037] Figure 6 This is a schematic diagram of the neck network connection of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example

[0040] like Figure 1 As shown in the flowchart, an embodiment of the invention provides a gradient-layered multi-branch air-to-air UAV target detection method, which specifically includes the following steps:

[0041] S1, Prepare the dataset: Prepare a suitable image dataset of air-to-air drone targets;

[0042] S1 collects video sequences captured by air-to-air drones, covering different lighting, weather, and occlusion scenarios to ensure the inclusion of multi-scale targets and dynamic motion scenes. Continuous frames are extracted from the video at a fixed frame rate, preserving the temporal correlation between adjacent frames, which serves as the source of spatial features for network input.

[0043] Secondly, to meet the optical flow input requirements of the spatiotemporal scale-aware ST-C2f feature extraction module, the optical flow field is calculated for each pair of adjacent frames (using models such as RAFT), and paired with the corresponding frame images for storage to form "image-optical flow" sample pairs. At the same time, the size of the image and the optical flow field are unified to match the network input resolution.

[0044] Then, a four-scale hierarchical annotation strategy is adopted when annotating the data: the bounding box accuracy is enhanced for ultra-small targets, the complete outline is annotated for medium and larger targets, and the target motion state and occlusion degree are recorded at the same time to provide a basis for dynamic weight adjustment in the loss function.

[0045] Finally, the dataset is divided into training, validation, and test sets (in a ratio of approximately 7:2:1) to ensure that the target scale, motion state, and scene distribution are consistent across the sets. Data augmentation is performed on the training set, including random cropping, scaling, and brightness / contrast adjustment, while maintaining consistency in augmentation operations for adjacent frames to avoid disrupting temporal relationships. This results in a spatiotemporal fusion dataset adapted for network training.

[0046] S2, Design a multi-branch downsampling module: such as Figure 4 The three-branch downsampling module contains three branches: Branch 1 is a standard convolution branch, which performs downsampling through a 3×3 convolution (stride 2) to capture local texture and edge features, followed by BN and SiLU activation; Branch 2 is a max-pooling branch, which first uses 3×3 max pooling (stride 2) to preserve significant spatial structure information, then uses a 1×1 convolution to adjust the number of channels, and finally uses BN and SiLU activation, with the output dimension consistent with Branch 1; Branch 3 uses a 3×3 depthwise separable convolution (stride 2) to reduce computation while preserving global spatial layout features, followed by a 1×1 pointwise convolution to compress channels, and outputs after BN and SiLU activation. The two-branch downsampling module contains two branches, retaining the standard convolution branch and the depthwise separable convolution branch (structure as above), removing the max-pooling branch to reduce redundant computation, and focusing on extracting semantic features of large targets.

[0047] like Figure 3Based on the characteristics of different resolution layers, the number and type of gradient branches are adjusted. A three-branch structure is used first in high-resolution layers and a two-branch structure is used in low-resolution layers. This approach enables comprehensive capture of multi-scale features and efficient allocation of computing resources, improving accuracy and making it suitable for air-to-air UAV detection applications.

[0048] S3, Design of the spatiotemporal-scale-aware ST-C2f feature extraction module: such as Figure 4 The spatiotemporal scale-aware ST-C2f feature extraction module first passes the input data through a spatiotemporal attention module before entering the first convolutional layer. Then, it is input to bottleneck module 3, and then residual connections are made with bottleneck modules 1, 2, and 3 respectively. The outputs from bottleneck modules 3 to 4 and from bottleneck modules 1 to 2 are concatenated and then output into the second convolutional layer. Bottleneck modules 1 and 2 are each composed of one convolutional layer, one deformable convolutional layer, and one convolutional layer, respectively. Bottleneck modules 3 and 4 are each composed of variable convolutional layers. This structure improves the feature extraction capability, enhances the accuracy of the object detection network, and significantly improves the adaptability to aerial scenes.

[0049] The spatiotemporal attention module integrates spatial and temporal information, compresses high-dimensional optical flow information and aligns it with the feature map, and then concatenates the original feature map with the preprocessed optical flow field in the channel dimension. It is suitable for dynamic target detection tasks in air-to-air scenarios, with efficient spatiotemporal information fusion and dynamic attention to moving regions. Spatial attention generates a spatial weight map by calculating the feature similarity of different positions within the same frame, which enhances the feature representation of key regions. Temporal attention generates a temporal weight sequence by comparing the feature differences between adjacent frames, which suppresses static background interference and highlights the dynamic features of moving targets.

[0050] The formula for time attention is:

[0051]

[0052] W represents the temporal attention weights of the t-th video frame, where softmax is the normalization function. t It is a learnable weight matrix for the input feature X at time t. t Perform a linear transformation, b t This is a bias term used to optimize the stability of linear transformations.

[0053] The formula for spatial attention is:

[0054]

[0055] W represents the attention weight at the nth spatial location, where softmax is the normalization function.s It is a learnable spatial weight matrix, representing the input channel feature X at the nth spatial location. n Perform a linear transformation, b s This is a bias term used to optimize the stability of linear transformations.

[0056] The combined weights are:

[0057]

[0058] Important target regions within frames and single frames are selected in both time and space dimensions, retaining the region of moving drones, suppressing interference from complex backgrounds, and enhancing the feature representation of dynamic small targets.

[0059] S4, Design the loss function: During training, the loss function consists of four parts: classification loss (L_cls), bounding box regression loss (L_box), target confidence loss (L_obj), and motion perception loss (L_motion).

[0060] The classification loss function adopts an improved version of Focal Loss, which alleviates class imbalance through dynamic weights. The bounding box regression loss function adopts a combination of GIoU and DIoU loss functions. The target confidence loss adopts a weighted binary cross-entropy. The motion perception loss includes two parts: optical flow consistency loss function and temporal feature difference loss function.

[0061] The formula for the composite loss function described in S4 is as follows:

[0062] L_total=L_cls+2L_box+L_obj+1.5L_motion

[0063] The classification loss function formula is:

[0064] L_cls=FocalLoss(p,y)=-α(1-p) γ log(p)-(1-α)p γ (1-y)log(1-p)

[0065] γ is the weight of the hard-to-distinguish control samples, p is the predicted class probability by the model, and α is the class balance factor used to compensate for class imbalance. α is larger for the positive class and smaller for the negative class. Here, the positive class is the drone, and the negative class is the background. A label y of 1 represents the positive class, and y of 0 represents the negative class. The loss terms in these cases are -α(1-p). γ log(p) and -(1-α)p γIn the air-to-air scenario, the class imbalance problem can be adjusted by using the local classification loss to make the model focus more on the positive class. Adjusting α = 0.75 can increase the proportion of positive class loss. The difference in prediction confidence can also be adjusted by γ. The p of small targets is low, (1-p). γ This will significantly amplify the loss, causing the model to learn small target features.

[0066] The formula for the bounding box regression loss function is:

[0067] L_box=GIoULoss+0.5DIoUAwareLoss

[0068] The bounding box regression loss function uses a combination of GIoU and DIoU loss functions.

[0069] The formula for the GIoU loss function is as follows:

[0070] GIoU=IoU-|C-(A∪B)| / |C|

[0071] GIoULoss=1-GIoU

[0072] A is the predicted UAV location bounding box output by the model, B is the ground truth bounding box, C is the smallest rectangle that contains both A and B, IoU is the intersection-union ratio, |C-(A∪B)| / |C| is the area of ​​the blank area that does not belong to the predicted bounding box or the ground truth bounding box divided by the area of ​​C. The optimization logic is to adjust the coordinates of the predicted bounding box through backpropagation to minimize the GIoU loss.

[0073] The DIoU loss function formula is as follows, where ρ is the distance to the center point and c is the length of the diagonal of the minimum closed box.

[0074] DIoU = IoU - p 2 (b,b g ) / c 2

[0075] DIoUAwareLoss=1-DIoU

[0076] IoU is the intersection-union ratio, p 2 (b,b g c is the squared Euclidean distance between the centers of the predicted bounding box and the ground truth bounding box. 2 It is the square of the diagonal length of the smallest bounding rectangle of the two. Based on IoU, the offset in the penalty is used to accelerate the regression of the drone frame, optimize the overlap and center position of the frame, and improve the positioning accuracy.

[0077] The bounding box regression loss uses a combination of GIoU and DIoU loss functions, which takes into account the convergence of small targets and the accuracy of dynamic target localization, and greatly improves the robustness of bounding box regression in air-to-air scenarios.

[0078] The formula for the target confidence loss function is:

[0079]

[0080] The target confidence loss uses binary cross-entropy loss, with the input being the predicted target existence probability p(y). i ) and real label y i When y i When y = 1, it means there is a goal, y i When the value is 0, there is no target, and the loss is averaged across all positions to allow the model to learn the pattern of target existence.

[0081] Motion perception loss includes the optical flow consistency loss function and the temporal feature difference loss function, as shown in the following formula:

[0082] L_motion = L_flow + 0.7L_temp

[0083] The formula for optical flow consistency loss is (L_flow):

[0084] L_flow=‖Flow_pred-Flow_gt‖2

[0085] Where Flow_pred is the optical flow predicted by the network, and Flow_gt is the reference optical flow.

[0086] The formula for temporal feature difference loss is (L_temp):

[0087] L_temp=||F_t-Warp(F_{t-1}),Flow_gt||2

[0088] Warp is an operation that transforms the features of the previous frame based on optical flow. When the drone moves from the position of frame t-1 to frame t, Warp(F_{t-1}) will move the drone features in t-1 to the position of frame t. If the model accurately predicts the drone features F_t in frame t, the difference between F_t and Warp(F_{t-1}) will be small. If there is a missed detection, the difference will be large. Therefore, constrained dynamic drone detection is performed to reduce missed detections.

[0089] This composite loss function supervises the consistency between optical flow prediction and temporal features, enhances the detection capability of moving targets, incorporates optical flow compensation into the bounding box regression loss to solve the localization ambiguity problem of moving targets, and achieves advantages such as multi-scale balance, gradient stability, and reduced missed detections.

[0090] S5, Construct and train the object detection network model: (e.g.) Figure 2The diagram illustrates a gradient-layered, multi-branch air-to-air UAV target detection network, comprising a backbone network, a neck network, and four detection heads. The specific structure of the backbone network is shown below. Figure 3 Specifically, it includes one convolutional downsampling module, two three-branch downsampling modules, two two-branch downsampling modules, three spatiotemporal scale-aware ST-C2f feature extraction modules, one C2f feature extraction module, and an SPPF spatial pyramid pooling module. The convolutional downsampling module consists of convolutional layers, activation functions, and batch normalization layers. The convolutional downsampling module is used to reduce the size of the feature maps. The multi-branch downsampling module, such as... Figure 4 As shown in the figure, the spatiotemporal scale-aware ST-C2f feature extraction module structure diagram is as follows: Figure 5 As shown; the neck network is as follows Figure 6 As shown, the system consists of an upsampling module, a spatiotemporal scale-aware ST-C2f feature extraction module, a C2f module, and a convolution module. First, the PAN-FPN architecture is extended upwards, utilizing spatial information from shallow feature maps to preserve the high-resolution spatial details needed for detecting small targets. Second, to add a fourth detection scale, a direct connection is made from the backbone p2 to the neck p2. Finally, a spatiotemporal scale-aware ST-C2f feature extraction module is embedded in the path to enhance feature correlation between adjacent frames and reduce missed detections. Four detection heads are used to detect ultra-small, small, medium, and large targets, respectively. A dedicated feature processing module is designed for each detection head. For the detection heads of small and ultra-small targets, a spatiotemporal scale-aware ST-C2f feature extraction module is added to enhance the detection of moving small targets. For the detection heads of medium and large targets, deformable convolution is added to handle the shape changes of large targets, and optical flow-assisted bounding box prediction is introduced to improve the localization accuracy of moving targets.

[0091] S6. Selecting Evaluation Metrics: Input the validation set of the dataset into the network model of this invention to obtain the optimal model results and model accuracy, thereby achieving target detection. The evaluation metrics are tailored to the core requirements of air-to-air UAV detection scenarios, focusing on four dimensions: overall performance, scale sensitivity, dynamic adaptability, and computational efficiency. The effectiveness of model validation improvements is compared and compared. Therefore, the evaluation metrics used are single-threshold average accuracy, multi-threshold average accuracy, scale-based average accuracy, moving target AP, number of parameters, computational cost, and FPS.

[0092] The overall accuracy data comparison is as follows: the mAP@0.5 of this invention is 8.6 percentage points higher than YOLOv8. The core advantage lies in the detection of ultra-small and small targets, thanks to the preservation of details by four detector heads and multi-branch downsampling. The AP for medium or large targets is close to that of v9 and v10, indicating that the improvement has not sacrificed the performance for large targets. The spatiotemporal fusion of the spatiotemporal scale-aware ST-C2f feature extraction module is effective for targets of multiple scales.

[0093]

[0094] Experimental results show that the AP of moving targets is significantly improved (14.3% higher than v8, 9.7% higher than v9, and 11.0% higher than v10). The optical flow fusion and temporal loss of the spatiotemporal scale-aware ST-C2f feature extraction module effectively enhance the dynamic target response.

[0095] Model Number of parameters (M) FLOPs(G) FPS YOLOv8 28.5 80.2 52 YOLOv9 32.1 95.6 45 YOLOv10 27.8 78.5 55 This invention network 29.3 83.7 47

[0096] The number of parameters increased by only 3.2% and FLOPs increased by 4.4%, which is far lower than the 12.6% increase in parameters and 19.2% increase in FLOPs in v9, demonstrating the lightweight design advantages of the multi-branch downsampling module and the spatiotemporal scale-aware ST-C2f feature extraction module.

[0097] The FPS remains at 47, meeting real-time requirements (greater than 30 FPS). Although slightly lower than v8 and v10, the improvement in accuracy far outweighs the efficiency loss, making it particularly suitable for accuracy-sensitive scenarios such as air-to-air drones.

[0098] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A gradient-layered multi-branch air-to-air unmanned aerial vehicle target detection method, characterized in that: The method specifically includes the following steps: S1, Prepare the dataset: Prepare an image dataset of air-to-air UAV targets, including images and extracted optical flow information; S2, Design of a multi-branch downsampling module: This module includes a three-branch downsampling module and a two-branch downsampling module, which contain standard convolution, deformable convolution, depthwise separable convolution, BN layer, max pooling and activation function; S3, Design of the Spatiotemporal Scale-Aware ST-C2f Feature Extraction Module: This module includes a spatiotemporal attention mechanism, standard convolution, deformable convolution, and the convolutional layers contain normalization and activation functions. Finally, they are connected and combined through a designed residual structure. The spatiotemporal scale-aware ST-C2f feature extraction module is composed of a spatiotemporal attention module and four bottleneck modules consisting of deformable convolution and ordinary convolution, which are residually connected. In addition, passing the spatiotemporal scale-aware ST-C2f feature extraction module through a C2f module and an SPPF module in high and medium resolution layers is the optimal strategy to balance performance and efficiency. S4, Design the loss function: Design a composite loss function, including classification loss, bounding box regression loss, target confidence loss and motion perception loss; S5, Construct and train the object detection network model: Construct a new object detection network by adding a multi-branch downsampling module and a spatiotemporal scale-aware ST-C2f feature extraction module. The object detection network model includes a backbone network, a neck network and four detection heads. S6, Select Evaluation Metrics: Select the best evaluation metrics to measure the accuracy of the gradient hierarchical multi-branch air-to-air UAV target detection method and evaluate the target detection results completed in S5.

2. The gradient-layered multi-branch air-to-air UAV target detection method according to claim 1, characterized in that: In S2, the multi-branch downsampling module adjusts the number and type of branches according to the characteristics of different resolution layers. First, a three-branch structure is used in the high-resolution layer, and a two-branch structure is used in the low-resolution layer. This can achieve comprehensive capture of multi-scale features and efficient allocation of computing resources, improve accuracy, and is suitable for air-to-air UAV detection application scenarios. The three-branch structure includes standard convolution max pooling and depthwise separable convolution, which are used to capture local features while preserving texture edge information and preserve spatial structure, respectively. The dual-branch structure includes standard convolution and depthwise separable convolution, which are used to capture semantic features and perform lightweight downsampling, respectively, balancing accuracy and efficiency.

3. The gradient-layered multi-branch air-to-air UAV target detection method according to claim 1, characterized in that: In S3, the spatiotemporal attention module compresses high-dimensional optical flow information and aligns it with the feature map by fusing spatial and temporal dimension information. Then, it splices the original feature map with the preprocessed optical flow field in the channel dimension. This is suitable for dynamic target detection tasks in air-to-air scenarios, with efficient spatiotemporal information fusion and dynamic attention to moving regions.

4. The gradient-layered multi-branch air-to-air UAV target detection method according to claim 1, characterized in that: In S4, the loss function during training includes four parts: classification loss, bounding box regression loss, target confidence loss, and motion perception loss. The classification loss function adopts an improved version of Focal Loss, which alleviates class imbalance through dynamic weights. The bounding box regression loss function adopts a combination of GIoU and DIoU loss functions. The target confidence loss adopts a weighted binary cross-entropy. The motion perception loss includes two parts: optical flow consistency loss function and temporal feature difference loss function. This loss function has the advantages of multi-scale adaptation, dynamic scene optimization, good module collaboration and improved robustness.

5. The gradient-layered multi-branch air-to-air UAV target detection method according to claim 1, characterized in that: In S5, the new target detection network backbone network goes through three spatiotemporal scale-aware ST-C2f feature extraction modules, one C2f module, one SPPF module, and a multi-branch downsampling module. Then, through the neck network, it undergoes top-down and bottom-up feature fusion and feature extraction to obtain a multi-scale information feature map. Finally, four detection heads are added. The four detection heads are used to detect large targets, medium targets, small targets, and micro targets, respectively.

Images