摘要
野生动物作为生态系统的重要组成部分,其动态监测对于维系生态平衡、理解物种间相互作用及评估生态系统健康状况具有至关重要的意义。野生动物监测主要通过无人机机载相机和固定的红外相机来捕捉动物的自然行为。然而,由于野生动物行为的不可预测性,在实际跟踪过程中,常会出现目标较小、多尺度变化以及动物身体被遮挡等问题。为了应对这些挑战,提出一种基于改进孪生网络的动物目标跟踪方法,将跟踪问题转化为相似性学习问题。在孪生关系网络(SiamRN)的特征提取阶段引入多头注意力机制,包括串联窗口自注意力运算和滑动窗口自注意力运算,增强模型对小目标的精准跟踪能力。同时,多头注意力机制的引入降低了网络的参数量和复杂度,提高了运算效率。在公开数据集和自制数据集上进行实验,结果表明本研究采用的野生动物跟踪方法的成功率和准确率分别为0.698和0.928,优于主流的孪生网络跟踪方法,该方法能够准确跟踪和定位野生动物目标,实现野生动物监测。
As an important component of the ecosystem,the dynamic monitoring of wildlife is of great significance for main⁃taining ecological balance,understanding species interactions,and assessing the health status of the ecosystem.Wildlife monitoring mainly relies on unmanned aerial vehicle(UAV)onboard cameras and fixed infrared cameras to capture the natural behavior of animals.However,due to the unpredictability of wildlife behavior,there are issues with small targets,multi-scale variations,and animal body occlusion in the actual tracking process.To address these challenges,this paper proposed an animal target tracking method based on an improved Siamese network,which transforms the tracking problem into a similarity learning problem.Introducing multi head attention mechanism in the feature extraction stage of Siamese re⁃lation network(SiamRN),including concatenated window self-attention operation and sliding window self-attention opera⁃tion,the precise tracking ability of small targets is enhanced.At the same time,the introduction of multi head attention mechanism reduces the number and complexity of network parameters and improves computational efficiency.The experi⁃ment was conducted on both public and self-made datasets,and the results showed that the success rate and accuracy of the wild animal tracking method used in this paper are 0.698 and 0.928,respectively,which was superior to mainstream Siamese network tracking methods.The method proposed in this paper can accurately track and locate wildlife targets,achieving wildlife monitoring.
作者
殷子璇
赵亚琴
肖治术
肖文宏
虞秋萍
许智皓
YIN Zixuan;ZHAO Yaqin;XIAO Zhishu;XIAO Wenhong;YU Qiuping;XU Zhihao(College of Mechanical and Electronic Engineering,Nanjing Forestry University,Nanjing 210037,China;State Key Laboratory of Integrated Management of Pest Insects and Rodents in Agriculture,Institute of Zoology,Chinese Academy of Sciences,Beijing 100101,China)
出处
《野生动物学报》
北大核心
2025年第3期533-543,共11页
CHINESE JOURNAL OF WILDLIFE
基金
国家自然科学基金面上项目(32371583)。
关键词
单目标跟踪
野生动物
孪生网络
注意力机制
Single target tracking
Wildlife
Siamese network
Attention mechanism
