基于人体骨架的扶梯乘客异常行为识别方法  

An abnormal behavior recognition method of escalator passengers based on human skeletons

作　　者：杨学存[1,2] 李杰华 陈丽媛 季韦 张尚辉 YANG Xuecun;LI Jiehua;CHEN Liyuan;JI Wei;ZHANG Shanghui(College of Electrical and Control Engineering,Xi'an University of Science and Technology,Xi'an 710054,China;Xi'an Key Laboratory of Eletrical Equipment Condition Monitoring and Power Supply Security,Xi'an 710054,China)

机构地区：[1]西安科技大学电气与控制工程学院,西安710054 [2]西安市电气设备状态监测与供电安全重点实验室,西安710054

出　　处：《安全与环境学报》2024年第2期636-643,共8页Journal of Safety and Environment

摘　　要：为准确识别乘客搭乘自动扶梯时的异常行为,避免安全事故的发生,提出了一种基于人体骨架的扶梯乘客异常行为识别方法。首先使用YOLOX-Tiny对视频中乘客位置进行检测,通过Alphapose算法提取骨骼关键点坐标,降低复杂背景的干扰;再使用多流膨胀3D卷积模块增强时空特征提取能力,聚合乘客骨架的全局特征;然后将其输入改进后的时空图卷积网络中提取乘客骨架信息,通过MS-TCN模块扩大接受域以增强时间特征的提取,联合人体关键点注意力模块(Key Point Attention Module,KPAM)提升网络对相似动作的关键骨架的关注度;最后通过Softmax对异常动作进行分类。采集扶梯运行现场视频制作数据集,试验结果表明,本文算法对乘客异常行为的识别精度达到96.1%,可应用于扶梯现场的视频监控系统,提高安全管理信息化水平。To accurately identify the abnormal behavior of passengers taking escalators,aiming at the problem that the traditional behavior recognition method is easily affected by the environment and the accuracy is low,this paper proposed an escalator passenger abnormal behavior recognition method based on the human skeleton.Firstly,the position of passengers in the input video was detected by YOLOX-Tiny.Then,the coordinates of key points of passengers'bones were extracted by the Alphapose algorithm to obtain the skeleton information in the video stream containing the actions of passengers in the process of taking the escalator.The skeleton diagram structure was constructed to obtain the initial human feature map and reduce the interference of complex backgrounds.Then,the multi-stream dilated 3D convolution module is used to enhance the ability of time feature extraction and motion information capture and aggregate the global features of the passenger skeleton,which not only solves the problem of insufficient utilization of skeletal map information in the Spatio-Temporal Graph Convolutional Network(ST GCN)network but also learns more discernable features.Then,the passenger skeleton information is extracted from the improved ST GCN.The MS TCN module is used to expand the acceptance field,enhance the extraction of temporal features,and improve the generalization ability of the network.The Key Point Attention Module(KPAM)was combined to enhance the network's attention to the key skeleton,reduce the misjudgment of similar actions in the process of action recognition and improve the recognition accuracy.Finally,the Softmax classifier is used to classify abnormal actions.The video collected from the scene of the escalator operation is cropped to make the dataset.The experimental results show that the recognition accuracy of the proposed algorithm for common abnormal behaviors is up to 96.10%,and it has a strong generalization ability.In addition,the proposed algorithm is compared with the LSTM network,GCN network,and the origin

关 键 词：安全工程 扶梯乘客异常行为 时空图卷积网络 人体骨架信息 行为识别 

分 类 号：X951[环境科学与工程—安全科学]