室内场景下人体行为异常检测方法研究  

Research on Detection of Abnormal Behavior of Human Indoors

作　　者：文峰[1] 刘飞[1] 黄海新[1] WEN Feng;LIU Fei;HUANG Haixin(Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学信息科学与工程学院,沈阳110159

出　　处：《沈阳理工大学学报》2022年第2期20-25,共6页Journal of Shenyang Ligong University

基　　金：辽宁省高等学校创新人才支持计划资助项目;沈阳理工大学科研创新团队建设计划资助项目(SYLUTD202105)。

摘　　要：基于图像的人体异常行为检测方法,当人数增多、遮挡等情况发生时,人体行为数据信息可靠性较低,检测精度不高、自适应性差,本文提出一种基于姿态特征的异常行为检测方法解决上述问题。利用成熟的人体姿态识别技术提取视频中人体关节点数据,将关节点坐标转化为人体行为的角度特征和距离特征以表达人体姿态;应用机器学习方法对关节特征进行分析和处理,获取有利于标识异常动作的数据分布特征;采用聚类算法在视频序列中对异常行为进行标记;使用支持向量机识别具体异常动作种类,实现人体异常行为检测。实验结果表明,相比于基于图像检测的方法,该算法检测精度达到了89.65%,可以运用于室内人体行为检测。Detection method of the image-based human abnormal behavior has low reliability,low detection accuracy and poor adaptability when the number of people increases and occlusion occurs.This paper proposes an abnormal behavior detection method based on attitude characteristics to solve the above problems.Mature human pose recognition technology was used to extract the human body node data in the video,and an Angle feature and distance feature were designed to express human body posture by transforming the coordinates of human body node into human behavior.The joint features were analyzed and processed by machine learning method to obtain the data distribution features which were helpful to identify abnormal movements.A clustering algorithm is used to mark abnormal behaviors in video sequences.Support vector machine is used to identify types of specific abnormal action and realize abnormal human behavior detection.Experimental results show that compared with the method based on image detection,the detection accuracy of this algorithm reaches 89.65%,which can be applied to indoor human behavior detection.

关 键 词：姿态识别 异常行为检测 聚类 支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]