手扶电梯(简称扶梯)乘客异常行为识别研究具有重要意义． 针对传统行为识别 算法易受环境影响、不能实时并准确对多目标进行识别的问题，提出一种基于人体骨架序 列的扶梯乘客异常行为识别算法． 该算法首先通过结合可变形组件模型特征的支持向量 机检测乘客人脸，并用改进的核相关滤波器对其进行跟踪，从而得到乘客在扶梯中的运动 轨迹;接着利用卷积神经网络提取轨迹中乘客的人体骨架序列，并通过模板匹配从乘客人 体骨架序列中检测异常行为骨架序列;最后利用动态时间规整将其与各类异常行为骨架 序列匹配，基于 k 近邻方法识别异常行为． 对 10 段扶梯视频的实验结果表明，文中所提的 异常行为识别算法处理速度达到 10 帧/秒，识别准确率为 93． 2%，能够实时、准确地识别 多种乘客异常行为．

The research on recognition of passengers＇ abnormal behavior on the escalator is of great significance． The traditional behavior recognition algorithm can not accurately recognize the multi-target in real time，and the re- cognition result is easily affected by the environment change． So an algorithm for recognizing passengers＇abnormal behavior on the escalator based on human skeleton sequences was proposed． Firstly，the passenger＇s face was de- tected by the support vector machine and tracked by the improved kernelized correlation filter to obtain the trajecto- ries of the passengers． Then，the human skeleton sequences of passengers were extracted by the convolutional neu- ral network． After that，the abnormal behavior skeleton sequences were detected from the human skeleton se- quences of passengers through template matching． Finally，the abnormal behavior was recognized by matching its skeleton sequence with all kinds of abnormal behavior skeleton sequences through dynamic time warping． The re- sults of experiment on 10 escalator videos show that the algorithm achieves a processing speed of 10 frames per se- cond and the recognition accuracy rate is 93. 2%，so it can accurately recognize a variety of passenger＇s abnormal behaviors in real time．