UAV-to-UAV Small Target Detection Method Based on Deep Learning in Complex Scenes

In the air-to-air unmanned aerial vehicle (UAV) detection scene, target UAVs often appear so small that they are difficult to detect due to the perspective changes and continuous motions of both the source and target UAVs in complex scenes. This article proposed a small target detection model for UAV (UAV-STD), which is supposed to improve the detection accuracy of small target UAVs in air-to-air scenes. The model integrates an attention mechanism-based small target detection (AMSTD) module, which effectively extracts and retains small target feature information of the UAV and improves focus on these features. Then, a spatial-aware and scale-aware prediction head (SSP Head) combining spatial perception and scale perception is constructed, which applies different attention in spatial and scale dimensions to improve the scale perception ability and spatial location perception ability of the model for small UAV targets. A bounding box loss function combining normalized Wasserstein distance and complete intersection over union (NWD-CIoU) is also proposed to improve the more accurate positioning of UAV small targets. Corresponding experimental results show that the average precision (AP) AP50 and APs of the UAV-STD model reach 83.7% and 83.1%, which are increased by 8.8% and 9.3%, respectively. Compared with other target detection methods, the proposed UAV-STD model performs well for small target UAVs in complex air-to-air scenes. This study provides excellent technical support for the application of UAVs in the field of safety monitoring and swarm confrontation. The code and model are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/LQS-IUST/UAV-STD</uri>.