铁路通信信号工程技术2025，Vol.22Issue(12)：1-11,11.DOI:10.3969/j.issn.1673-4440.2025.12.001
基于动态分布回归细化的航拍目标检测算法
Aerial Object Detection Algorithm Based on Dynamic Distribution Regression Refinement
Tao Zhilai 1Dai Rong 1Liu Chang 2Chen Yi2
作者信息
- 1. Hefei Metro Group Co.,Ltd.,Hefei 230000,China
- 2. CRSC Urban Rail Transit Technology Co.,Ltd.,Beijing 100070,China
- 折叠
摘要
Abstract
Addressing the challenges of small target size,scarce positive samples and localization uncertainty in vehicle and personnel object detection during Unmanned Aerial Vehicle(UAV)inspection,this paper proposes a Dynamic Distribution Regression Refinement(DDRR)algorithm.The core innovations of the proposed algorithm are as follows:An adaptive a priori movement module is designed,which utilizes adaptive temperature parameters to optimize anchor box-target matching quality.A coarse positive sample screening mechanism based on Distance Intersection over Union(DIoU)is established,which employs distance-aware geometric matching metrics to reduce computational complexity.A refined positive sample screening module based on Bhattacharyya distance is developed,which combines intelligent iterative strategies to strike a balance between computational efficiency and detection accuracy.Experiments on VisDrone and AI-TOD-V2 datasets demonstrate that the DDRR algorithm achieves Average Precision(AP)scores of 33.25%and 27.3%,showing significant improvements over mainstream detectors,with an inference speed of 19.8 Frames Per Second(FPS)meeting real-time requirements.The proposed algorithm systematically addresses a priori mismatch,sample imbalance and localization uncertainty issues in UAV inspection,providing an efficient and accurate detection solution for inspection systems.关键词
低空无人机巡检/动态分布回归/车辆人员检测/自适应先验移动/智能迭代优化Key words
low-altitude UAV inspection/dynamic distribution regression/vehicle and personnel detection/adaptive priori-movement/intelligent iterative optimization分类
航空航天引用本文复制引用
Tao Zhilai,Dai Rong,Liu Chang,Chen Yi..基于动态分布回归细化的航拍目标检测算法[J].铁路通信信号工程技术,2025,22(12):1-11,11.
