NST−YOLO: Improved YOLOv10 model for small target UAV detection

Abstract

This paper proposes a UAV small target detection algorithm based on improved YOLOv10, aiming to address the challenge of detecting small and elusive UAV targets in complex scenes. The model incorporates a hierarchical attention mechanism and neighborhood weighted distance loss (NWDLoss) of Swin Transformer, enhancing its capability to detect small targets. The Swin Transformer enhances feature extraction through its global context perception ability, while NWDLoss optimizes the geometric modeling process of bounding box regression, especially in dense scenes and small target detection. Experimental results demonstrate that the NST-YOLO model achieves an average accuracy of 69.89 % on a test set consisting of real objects and multiple videos; the mean Average Precision at IoU threshold 0.5 (mAP50) and across thresholds 0.5–0.95 (mAP50:95) of the enhanced model reach 94.02 % and 54.74 %, respectively. In addition, the model achieves 33.87 Frames Per Second (FPS) on the test system, which satisfies the real-time requirements of target recognition. By improving the backbone network and loss function, while ensuring real-time performance, it has better recognition and tracking effects than other algorithms. The enhanced model improves detection accuracy while maintaining real-time performance, particularly in low-light and small-target scenarios. Overall, the algorithm achieves a favorable balance between recognition accuracy and inference speed, and demonstrates robust performance against interference. These characteristics highlight its theoretical significance and practical potential for real-time detection and edge AI applications.