Deep Learning-Based Abnormal Elevator Action Detection System

In modern cities, elevators provide convenience and efficiency for people, but due to their small and enclosed environment, some abnormal actions may not be detected in a timely manner, resulting in safety accidents. In this paper, we analyze the abnormal actions that pedestrians may exhibit in the confined space of elevator cabins, and propose a complete end-to-end edge computing system for capturing passengers' actions from monitoring video streams accurately. The system filters out useless monitoring videos using YOLO algorithm, extracts features using the lightweight Mobilenet network, and finally recognizes actions using the Bi-LSTM network with attention mechanism. To address the problem of the lack of relevant datasets in the elevator environment, we propose a new self-made dataset. Our method has a low computational cost and high detection accuracy, achieving an accuracy of 95.89% on the dataset.