Detection And Research Of Dangerous Behavior Based On Video Monitoring Of Elevator

Massive amounts of shopping malls,rail transit and other public places in the city have installed escalators as an important tool to transport pedestrians.The escalator can effectively solve the problem of time-consuming and laborious that people up and down the escalator the elevator which becomes an indispensable part of people’s life.However,because the escalator has the characteristics of long-term uninterrupted running,the escalator cannot be braked in time which will cause continuous injury to personal safety once a dangerous situation occurs.The traditional method solves this problem by manual monitoring or early warning of the escalator,which not only has the problems of labor cost and low automation,but also has a slow response speed and cannot perform full-time monitoring.It cannot be applied widely.This article applies artificial intelligence and computer vision technologies to the field of safe-running escalators so as to solve the above problems.The algorithm captures the real-time video scene of the escalator through the camera,and then detects pedestrian behavior in the video scene,so as to give early warning when a safety accident occurs.This paper mainly detects these two dangerous behaviors of pedestrian fall and pedestrian retrograde.(1)The detection method based on multi-object tracking algorithm of pedestrian retrograde behavior is introduced.The method is based on the multi-target tracking algorithm Deep Sort combined with specific steps to complete retrograde detection.First,according to the tracking algorithm to obtain the pedestrian’s trajectory.Then make a judgment on the movement state of the pedestrian according to the retrograde detection step.Because of the Deep Sort algorithm is a track-by-detection idea to track multiple pedestrians in the video,so the method proposed in the article detects the target through the target detection algorithm.Then call the tracking module to obtain the pedestrian trajectory.Finally,the direction of each pedestrian trajectory is obtained by analyzing the motion trajectory.Pedestrian status can be obtained by comparing with the set running direction.The experiment on the self-built data set PRV-2019 verifies which the detection method detects pedestrian retrograde behavior excellently on the escalator scene and issues early warning information.(2)A pedestrian fall detection based on human body structure algorithm is proposed.Firstly,the Open Pose algorithm is used to get the coordinates of body keypoint in video.Then the coordinates is feature processing by the module.Finally,the classifier is used to classify features and detect whether there is a pedestrian fall according to the classification results.For fall behavior in escalator scene,this thesis constructs a pedestrian fall dataset called PFV-2019.It is proved through experiments that the real-time and effectiveness of the detection method on dataset.