Microscale Perception and Dynamic Label Assignment Object Detection Network for UAV Visual Sensing Systems

Abstract

To address the annihilation of high-frequency features and the dense overlapping of small targets captured by uncrewed aerial vehicle (UAV)-borne visual sensors in complex high-altitude environments, this article proposes PCS-you only look once (YOLO), an end-to-end sensing data processing and physical signal optimization network. First, a microscale perception reconstruction mechanism is designed. By constructing a high-resolution P2 branch coupled with the content-aware physical feature reassembly operator (CARAFE), weak physical features are extracted from the bottom architecture for sensing feature alignment and spatial resampling, effectively overcoming spatial faults and signal aliasing. Second, to resolve bounding box matching conflicts triggered by weak features in dense sensor clutter, the simplified optimal transport assignment (SimOTA) dynamic label assignment strategy is introduced. It adaptively assigns optimal perception anchor boxes for elusive tiny targets based on global optimal transport. Experiments on the VisDrone2019 sensing dataset demonstrate that the improved algorithm achieves 49.4% precision and 42.4% recall. The mean average precision (mAP)@0.5 and mAP@0.5:0.95 reach 41.8% and 23.3%, outperforming the baseline YOLOv5s by 7.9% and 4.8%, respectively. As an efficient software compensation scheme, PCS-YOLO overcomes hardware optical resolution limits while ensuring the real-time performance of UAV edge sensing nodes, significantly enhancing the detection precision of dense, tiny physical targets.