复杂场景下基于改进的YOLOv5-pose的异常行为检测研究  

Research on Abnormal Behavior Detection Based on Improved YOLOv5-pose in Complex Scenes

作　　者：崔悦 杨旺 王明旗 李振鑫 CUI Yue;YANG Wang;WANG Mingqi;LI Zhenxin(School of Information Engineering,Xinjiang Institute of Technology,Aksu 843100,China)

机构地区：[1]新疆理工学院信息工程学院,新疆阿克苏843100

出　　处：《现代信息科技》2025年第7期71-75,82,共6页Modern Information Technology

摘　　要：文章提出了一种复杂场景下基于改进的YOLOv5-pose的异常行为检测算法,使用FPT代替了FPN+PAN模块,使特征图能够在跨尺度与跨空间实现全局部交互,提高关节点检测的准确度。在Neck模块中使用跳跃连接结构将输入特征和经过网络输出的多尺度特征的信息进行有效融合,提高对细节信息的捕获能力,增强对遮挡过的关节点检测的准确性。实验结果表明,改进后的算法在CrowdPose数据集上平均准确率达到了99.5%,比原模型高出了2.4%。改进后的模型不仅具有更高的检测精度,而且在小目标的识别效果上也有显著的提升。This paper proposes an abnormal behavior detection algorithm based on the improved YOLOv5-pose in complex scenes.It uses FPT to replace the FPN+PAN module,enabling the feature maps to achieve global and local interaction across scales and spaces,and improving the accuracy of joint point detection.In the Neck module,a skip connection structure is employed to effectively fuse the information of the input features and the multi-scale features output through the network,improving the ability to capture detailed information and enhancing the accuracy of detecting occluded joint points.Experimental results show that the improved algorithm achieves an average accuracy of 99.5%on the CrowdPose dataset,which is 2.4%higher than that of the original model.The improved model not only has higher detection accuracy,but also significantly improves the recognition performance of small targets.

关 键 词：YOLOv5-pose 行为识别 关节点检测 FPT 跳跃连接 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]