基于YOLOv7-Tiny算法的无人机实时跟踪野生动物方法

Real Time Tracking of Wild Animals Using Unmanned Aerial Vehicles Based on YOLOv7-Tiny Algorithm

借助无人机边缘计算技术监测野生动物的运动状态和种群发展变化已成为科研工作者广泛使用的技术手段。传统跟踪算法算力高,机载边缘设备硬件资源算力不足,在户外复杂的自然环境下难以实现实时跟踪。为解决野外环境中无人机跟踪野生动物时遇到树木遮挡和背景干扰导致无法准确实时跟踪的问题,选取东北地区东北虎(Panthera tigris altaica)、狍(Capreolus pygargus mantschuricus)和驯鹿(Rangifer tarandus phylarchus)为研究对象,以YOLOv7-Tiny+Bot-SORT作为检测跟踪的基础框架,提出了一种轻量化的无人机跟踪算法。首先,采用FasterNet网络减少模型冗余计算,增强特征图中目标区域关注度;其次,采用高效通道注意力机制实现局部跨通道交流,降低复杂环境对检测网络的影响,提升网络检测能力;最后,为降低计算成本,替换重识别网络,提高无人机跟踪速度。结果显示:提出的实时跟踪方法准确度(MOTA)和精确度(MOTP)分别达到79.93%和73.48%,跟踪速度从3.4帧/s提升到43.4帧/s。研究表明,提出的算法不仅在提升跟踪精度和速度方面表现出色,而且更适用于算力有限的边缘设备,为保护野生动物的多样性和群体行为研究提供了强大的技术支持。

Utilizing UAV edge computing technology to monitor the movement pattern and population dynamic of wild animals has become a widely adopted technique among researchers.Traditional tracking algorithms demand high computational power,and pose challenges for onboard edge device with limited hardware resources,especially in complex outdoor environment where real-time tracking is difficult to achieve.To address issues such as tree cover and background interference that hinder accurate real-time tracking during UAV-based wild animals monitoring in outdoor settings,we selected northeast China’s wild animals species,including the Amur tiger(Panthera tigris altaica),roe deer(Capreolus pygargus mantschuricus),and reindeer(Rangifer tarandus phylarchus),as research subjects.Leveraging YOLOv7-Tiny+Bot-SORT as the detection and tracking framework,we proposed a lightweight UAV tracking algorithm.Initially,we employed the FasterNet network to reduce redundant computations in the model and enhance focus on target regions in feature maps.Subsequently,an efficient channel attention mechanism facilitated local inter-channel communication to mitigate the impact of complex environments on detection networks,thereby improving detection capability.Finally,to lower computational costs,we substituted the re-identification network to enhance UAV tracking speed.Results revealed that the proposed real-time tracking method achieved accuracy rates of 79.93% for multi-object tracking accuracy(MOTA)and 73.48% for multi-object tracking precision(MOTP),with tracking speed increasing from 3.4 to 43.4 frames per second.These findings demonstrate not only excellent performance in enhancing tracking accuracy and speed but also suitability for edge devices with limited computational resources,thereby providing robust technical support for biodiversity conservation and population behavioral studies.