Adaptive Feature Fusion and Improved Attention Mechanism-Based Small Object Detection for UAV Target Tracking

With the development of artificial intelligence technology, UAVs have the ability to perceive the environment. UAV combined with target detection technology for road environment perception has received extensive attention. However, the complexity and variability of real-world road environments pose challenges for target detection. To address these challenges, we propose a UAV small target detection algorithm AS-YOLOV5 with adaptive feature fusion and improved attention mechanism. In the feature extraction phase, AS-YOLOV5 employs soft pooling to bolster the feature extraction network, mitigating the loss of critical edge information of small targets inherent in standard down-sampling techniques. Our feature fusion method incorporates learnable parameters, effectively rebalancing feature layers to ensure small target information remains significant during the fusion process and is not obscured by large target features. The subspace attention module is optimized to enhance the representation of small target features while suppressing background interference. To ensure the detection branch captures essential shallow information about small objects, we introduce an additional feature extraction layer for feature fusion. Simulation results show that the proposed method outperforms the existing algorithms. AS-YOLOV5 achieved a mean average precision (mAP) of 56.36% on the BDD100K data set and an impressive 93.33% on the KITTI data set with an Intersection over Union (IoU) threshold of 0.5.