Abstract: Aiming at the problem that dangerous behavior of passengers on the escalator is difficult to be accurately detected in real time，an algorithm for identifying abnormal behavior of escalator passengers based on video surveil-lance was proposed． Firstly，YOLOv3 was used to detect the position of the passenger in the image． Secondly，MobileNetv2 was used as the base network，which was combined with the deconvolution layer to extract the human skeleton of the detected passenger． Thirdly，the Hungarian assignment algorithm based on skeleton distance was used to realize the allocation of passenger ID numbers in the video． Finally，with keypoints as the input，the graph convolutional neural network was used to recognize the abnormal behavior of passenger． The experimental results on the GTX1080GPU show that the proposed recognition algorithm can achieve a processing speed of 15 f/s and an abnormal behavior recognition accuracy rate of 94. 3%，which can accurately recognize the abnormal behavior of passengers on the escalator in real time．

Key words: escalator, deep learning, convolutional neural networks, pedestrian detection, pose estimation, Hungarian assignment algorithm