Air-to-Air Target Detection of Unmanned Aerial Vehicles Under High Dynamic Scenarios

Abstract

To address the issues of limited onboard computing resources and the difficulty of recognizing small targets in Unmanned Aerial Vehicle (UAV) air-to-air target detection tasks in high dynamic scenarios, a UAV air-to-air target detection algorithm based on a lightweight attention mechanism, SGC-YOLOv5, is proposed. First, the S-Ghost module and SD-Ghost structure are used to build an SD-GhostNet, a feature extraction network, which reduces the computational complexity and number of parameters of the model. Second, more efficient GSConv and VOVGSCSP structures are introduced to refine the feature fusion network, and SD-GhostNet is combined with the refined feature fusion network to achieve the best lightweight effect of the mode. Finally, a lightweight Convolutional Block Attention Module (CBAM) is added to the feature fusion network to highlight the UAV features of interest in the image, suppress redundant background information, and improve detection accuracy. Experimental results on the Det-Fly dataset indicate that the SGC-YOLOv5 algorithm achieves an accuracy, parameter count, detection speed, Floating Point Operations Per Second (FLOPs) of 74.9%, 4 313 695, 169.42 frames per second, and 9.0×10⁹, respectively. Compared with the benchmark YOLOv5s algorithm, the detection accuracy and detection speed are improved by 2.5% and 26.17 frames per second, respectively, whereas the parameter count and FLOPs are reduced by 48.5% and 57.5%, respectively. This model achieves good detection accuracy while achieving a lightweight model.