Improved YOLOv5 algorithm in quadcopter UAV target detection

The rapid advancement of Unmanned Aerial Vehicles (UAVs), intelligent monitoring systems, and related technologies has heightened the need for efficient small target detection, emphasizing the importance of developing more effective algorithms. To address the difficulties in accurately detecting small and medium-sized targets in video, this paper introduces an algorithm specifically for detecting slow-moving small targets. Traditional image processing techniques are used for identifying slow-moving small targets. The paper proposes a deep learning-based algorithm that enhances feature extraction by incorporating a CBAM attention module into the YOLOv5 backbone network. The BiFPN structure replaces the PANet structure in the neck network to better utilize underlying features. Additionally, a high-resolution detection head is integrated into the detection head to improve the ability to detect small, slow-moving targets. The algorithm is trained on the VisDrone2019 UAV dataset, and experimental comparisons demonstrate that the proposed slow-moving small target detection algorithm outperforms the standard YOLOv5, achieving a 13% improvement in mAP for small target detection in video.