摘要
红外相机是监测野生动物的常用方法,但存在数据量大、背景信息复杂等问题,导致监测数据标注和检测困难。针对以上问题提出一种基于伪标签和YOLOv4的野生动物检测方法。本方法首先基于运动检测的伪标签标定方法,通过背景差分法和形态学操作实现对视频数据集的自动快速标注,降低监测环境中复杂背景的不利影响;然后通过跨阶段局部卷积块,减少YOLOv4中路径聚合网络所需的计算量;最后在密集卷积区域引入Swish激活函数,提高模型在深层区域的特征提取能力,以浙江江山仙霞岭自然保护区的6种野生动物监测视频作为数据集进行实验。结果表明:本方法在平均精度均值和帧率指标上达到了86.41%和18.93帧/s,相比于YOLOv4、RFCN、YOLOv8x算法分别提高1.62、3.43、7.11个百分点,证明所提出算法可以有效克服现有方法标注和检测困难的问题,提升了野生动物监测数据检测平均精度均值和帧率,有助于野生动物监测数据分析的自动化和智能化。
The camera traps is a common method to monitor wildlife.A large amount of data and complex background in⁃formation make it difficult to label and detect.In order to solve the above problems,a wild animal detection method based on pseudo-labels and YOLOv4 was proposed.Firstly,a pseudo-labeling calibration method based on motion detection was proposed to realize automatic and fast calibration of video datasets by background difference method and morphological operation,which solved the problem of difficult automatic labeling caused by complex background in monitoring data.Then,a cross stage partial block was proposed to reduce the amount of computation required for the path aggregation net⁃work in YOLOv4.Finally,the Swish activation function was introduced in the dense convolution area,which improves the feature extraction ability of the model in the deep region.In this study,the monitoring videos of six kinds of wild animals in Zhejiang Jiangshan Xianxialing Provincial Nature Reserve were used as data sets for experiments.The results show that the method proposed in this study achieves 86.41%mean average precision(mAP)and a frame rate of 18.93 frames per second.These metrics represent improvements of 1.62 percentage points,3.43 percentage points,and 7.11 percentage points over the YOLOv4,RFCN,and YOLOv8x algorithms,respectively.This demonstrates that the proposed algorithm ef⁃fectively addresses the labeling and detection challenges faced by existing methods.Balanced improvement of detection mAP and frame rate contributes to automated and intelligent analysis of wildlife monitoring data.
作者
孟继森
马玉明
杨紫合
孙茜
巨友娟
谢将剑
张军国
MENG Jisen;MA Yuming;YANG Zihe;SUN Qian;JU Youjuan;XIE Jiangjian;ZHANG Junguo(Forestry Affairs Center of Tianjin Municipal Bureau of Planning and Natural Resources,Tianjin 300191,China;School of Technology,Beijing Forestry University,Beijing 100083,China;School of Information Engineering,Tianjin University of Commerce,Tianjin 300134,China;Forest Seed and Seeding Station of Huzhu Tu Autonomous County,Huzhu Tu Autonomous County 810500,China)
出处
《野生动物学报》
北大核心
2025年第3期523-532,共10页
CHINESE JOURNAL OF WILDLIFE
基金
中央财政林业科技推广示范项目(津〔2021〕JTG02号)
国家自然科学基金项目(32371874)。
关键词
野生动物
目标检测
红外相机
运动检测
伪标签
Wildlife
Object detection
Camera-trapping
Motion detection
Pseudo-labels
