基于MR-VOD的神农架林区野生动物视频检测  

A Multi-relation Video Object Detection of Wildlife in Shennongjia Forest Area

作　　者：眭海刚[1] 魏天怡 胡烈云 杨敬元 马国飞 SUI Haigang;WEI Tianyi;HU Lieyun;YANG Jingyuan;MA Guofei(State Key Laboratory of Information Engineering in Surveying,Wuhan University,Wuhan,430079,China;Shennongjia National Park Administration,Shennongjia,442421,China;Hubei Key Laboratory of Conservation Biology of Shennongjia Golden Monkey,Shennongjia,442421,China)

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,武汉430079 [2]神农架国家公园管理局,神农架442421 [3]神农架金丝猴保育生物学湖北省重点实验室,神农架442421

出　　处：《野生动物学报》2025年第1期1-13,共13页CHINESE JOURNAL OF WILDLIFE

基　　金：神农架国家公园本底资源综合调查研究项目(SNJNP2023015);神农架金丝猴保育生物学湖北省重点实验室开放课题基金项目(SNJGKL2023015)。

摘　　要：红外相机视频是野生动物调查的主流研究方法之一,但在林区受设备角度、复杂环境与野生动物活动的随机性影响,极易出现检测中光流捕捉错误或简单语义难以识别的情况。针对此类问题,提出一种基于多帧关系网络特征增强的视频目标检测方法(multi-relation video object detection,MR-VOD)。该算法在图像关系网络的基础上,综合考虑上下帧目标之间的关系,通过多阶段推理,实现对野生动物目标的准确检测。同时,以神农架林区野生动物红外相机视频为基础,构建相关野生动物视频目标检测数据样本集作为实验区。试验证明,改进后的算法检测性能有所提升,平均准确率达81.96%,比Faster R-CNN提高9.32个百分点,在川金丝猴(Rhinopithecus roxellana)的检测上提升30.79个百分点,并在多种复杂场景下测试表现良好,有效减少了错检漏检的情况。该算法的实现将为神农架野生动物智能监测云平台提供检测基础,同时为后续开展的野生动物保护、种群评估提供必要的技术支撑。Infrared camera video is one of the mainly research methods for wildlife investigation.However,due to the influ⁃ence of equipment angle,complex environment and randomness of wildlife activities in the forest area,it is very easy to be mistakenly captured with optical flow or hard to be recognized by simple semantics in the object detection.Thus,a multirelation video object detection(MR-VOD)based on multi-frame relational network feature enhancement was proposed.The algorithm,based on the image relation network,comprehensively considers the relationship between context targets and improve the accuracy of wildlife through multi-stage inference.Meanwhile,a Shennongjia wildlife video object detection da⁃taset extracted from infrared camera videos was built as the experimental area.The research proved that MR-VOD algo⁃rithm had improved detection performance,with an average accuracy of 81.96%,which is 9.32 percentage points higher than that of Faster R-CNN,and a significant improvement of 30.79 percentage points in the detection of golden sunb-nosed monkey(Rhinopithecus roxellana).Besides,the method has performed well in a serial of complex scenarios and reduces the mistakes and omissions during the detection.The implementation of the algorithm will provide a detection basis for the Shennongjia wildlife intelligent monitoring cloud platform,with the necessary technical support for the wildlife protection and population assessment subsequently.

关 键 词：关系网络 视频目标检测 野生动物 复杂环境 特征增强 

分 类 号：Q958.1[生物学—动物学] TP391.4[自动化与计算机技术—计算机应用技术]