摘要
东北虎(Panthera tigris altaica)是国家一级重点保护野生动物,在人工饲养环境下,对东北虎幼崽及其母虎活动的精准跟踪是研究东北虎幼崽成长过程中行为发育、观测东北虎个体健康状况的重要手段。本研究提出了一种轻量型的GhostNet-DeepSORT算法来实现在监控视频下的东北虎幼崽群体活动的多目标跟踪。该算法的测试结果表明:东北虎幼崽个体目标检测的召回率和平均精度均值分别是94.9%和96.2%,东北虎幼崽多目标跟踪准确度(MOTA)和精确度(MOTP)分别是91.6%和88.2%。通过轻量化操作处理后的GhostNet-DeepSORT算法,在MOTA保持不变时,MOTP提升了1.25%,而且模型占用内存从45.4 MB减小到6.5 MB。因此,相对于DeepSORT算法而言,GhostNet-DeepSORT算法在保证跟踪精度的同时更适用于算力资源不足的小型设备,相比于由宽残差网络构成的重识别网络,采用GhostNet网络进行模型的轻量化替代,实现在视频中东北虎个体重识别。该跟踪算法的实现也为后续圈养东北虎幼崽个体行为识别和个体健康的便捷和快速评估研究提供必要的技术支撑。
Amur tiger(Panthera tigris altaica)is one of the first-grade state key protected animals.The accurate tracking of the Amur tiger cubs and their mother in the artificial rearing and breeding environment is the basis for analyzing the individual behavior of the Amur tiger cubs during in the growing period and for observing health condition of the Amur tiger.In this study,a lightweight GhostNet-DeepSort network is proposed to realize multi-target tracking of juvenile Amur tigers group activities in surveillance video.According to the test results of the algorithm,the recall rate and average accuracy of the Amur tiger target detection are 94.9%and 96.2%respectively,and the tracking accuracy of the Amur tiger target MOTA and MOTP are 91.6%and 88.2%respectively.MOTP increases by 1.25%while MOTA remaines unchanged by GhostNet-DeepSORT algorithm after the lightweight operation,and the memory footprint of the model decreases from 45.4 MB to 6.5 MB.Therefore,with respect to DeepSORT algorithm,GhostNet-DeepSORT algorithm can ensure the tracking accuracy and is more suitable for small equipment with insufficient computing resources.Compared with the re-identification network composed of wide residual network,GhostNet network is used to replace the model with light weight to re-identify Amur tigers in video.The implementation of this tracking algorithm also provides necessary technical support for subsequent study on the convenient and rapid evaluation of individual behavior recognition and individual health of captive Amur tiger cubs.
作者
马光凯
吴伟
刘丹
崔永璐
邓雯心
姜广顺
MA Guangkai;WU Wei;LIU Dan;CUI Yonglu;DENG Wenxin;JIANG Guangshun(College of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin,150040,China;Feline Research Center of National Forestry and Grassland Administration,College of Wildlife and Protected Area,Northeast Forestry University,Harbin,150040,China;Heilongjiang Siberian Tiger Park,Harbin,150028,China)
出处
《野生动物学报》
北大核心
2022年第3期605-613,共9页
CHINESE JOURNAL OF WILDLIFE
基金
国家自然科学基金项目(31872241,32171691)
高等学校学科创新引智计划资助(B20088)
中央高校基本科研业务费专项资金项目(2572021BF08,2572019CP17)。
