基于检测和重识别的无人机行人跟踪算法  

UAV pedestrian tracking algorithm based on detection and re-identification

作　　者：张嘉辉 赵威 王子琛 蒙志君[1] ZHANG Jiahui;ZHAO Wei;WANG Zichen;MENG Zhijun(School of Aeronautic Science and Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]北京航空航天大学航空科学与工程学院,北京100191

出　　处：《北京航空航天大学学报》2024年第8期2538-2546,共9页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金(61976014)。

摘　　要：将智能检测跟踪算法与无人机(UAV)的灵活性相结合是UAV应用的研究热点。针对UAV的视角及运动导致目标滑移和遮挡的问题,提出一种基于检测和重识别的UAV行人跟踪算法。对训练好的YOLOv5进行TensorRT加速,解决UAV计算资源有限的问题;以量化加速的目标检测算法与重识别算法为基础,构建行人跟踪算法框架;设计判定行人匹配度,完成行人匹配系统设计。仿真试验表明:训练后的YOLOv5和OSNet具备一定的精度,采用TensorRT加速后的YOLOv5网络在保证精度的情况下,帧率有了近50%的提升。飞行试验表明:所提算法在行人穿插及障碍物遮挡的情况下,可以实现对目标的稳定跟踪,具备一定的实用性和有效性。Combining intelligent detection and tracking algorithms with the flexibility of unmanned aerial vehicle(UAV)is a hot research topic for UAV applications.A UAV pedestrian tracking algorithm based on detection and re-identification was proposed for solving the problems of target slippage and occlusion due to the UAV’s viewpoint and motion.Firstly,TensorRT acceleration of trained YOLOv5 was performed to solve the problem of limited UAV computational resources;secondly,a pedestrian tracking algorithm framework was constructed based on a target detection algorithm and a re-identification algorithm with quantization acceleration;finally,the pedestrian matching degree was designed and determined to complete the pedestrian matching system design.Simulation experiments show that the trained YOLOv5 and OSNet have certain accuracy,and the YOLOv5 network with TensorRT acceleration has nearly 50%improvement in frame rate with guaranteed accuracy.The flight test shows that the proposed algorithm can achieve stable tracking of the target under the situation of pedestrian intersection and obstacle occlusion,and it has certain practicality and effectiveness.

关 键 词：无人机 目标检测 量化加速 行人跟踪 重识别 

分 类 号：V279.2[航空宇航科学与技术—飞行器设计]