Research On Escalator Passenger Abnormal Behavior Detection Method Based On Machine Vision

With the rapid development of China’s economy and the acceleration of urbanization,more and more escalators are widely used in various public places.As a means of public transportation,escalator not only brings convenience to people,but also brings security risks that cannot be ignored,such as personal safety accidents caused by abnormal behaviors of passengers in the normal operation of escalator.In order to avoid the occurrence of the above-mentioned safety accidents,it is necessary to monitor the escalator area in real time.The traditional manual video monitoring method has problems that the monitoring personnel are easy to visual fatigue and cannot deal with emergencies in a timely manner.In order to solve the above problem,through to the escalator passengers abnormal behavior detection of an investigation of the current situation of the existing research,this paper puts forward a kind of escalator passengers,abnormal behavior detection based on machine vision framework,and puts forward an adaptive pruning algorithm of part classification network model of the detection algorithm with high efficiency and low loss compression.It can realize efficient detection and recognition of escalator passengers’ abnormal behavior,which has practical engineering value and important social significance.The main work contents are as follows:(1)The escalator passenger abnormal behavior data set EAB was established.This paper takes the abnormal behaviors of falling forward and climbing handrail belt,which are harmful to passengers,as the research objects.In the actual escalators of shopping malls,videos of three kinds of passenger behaviors including normal straight walking,falling forward and climbing the handrail belt were collected.The videos included 1,2 and 3 passengers for each kind of passenger behavior.Furthermore,the video was further processed by cutting frames,cutting and coding,and the original photo sets of escalator passengers’ behaviors of normal straight walking,falling forward and climbing handrail belt were established.Labelme was used to annotate these original image sets,further forming escalator passenger target detection dataset EAB-D,attitude estimation dataset EAB-P and abnormal behavior classification dataset EAB-C.(2)An escalator passenger abnormal behavior detection framework based on Alpha Pose is proposed.Firstly,the input image is detected by Yolov5,a target detection model with good accuracy and real-time performance,and then the frame is traced.Then,the top-down multi-person pose estimation algorithm Alpha Pose was used to frame the pedestrian node information and form the escalator passenger skeleton feature graph to improve the generalization performance of the classification model.Finally,the skeleton feature graph was classified by deep convolutional neural network classification model Efficient Net.In this paper,we train and test the network models of target detection,multi-person attitude estimation and skeleton feature graph classification of the proposed algorithm on EAB,a self-built escalator passenger abnormal behavior dataset.The experimental results show that the accuracy rate,the total average accuracy and the total average accuracy of escalator passengers’ abnormal behavior classification recognition by using the proposed detection framework can reach 87.9%,90.7% and 91.3% respectively.Under the condition of NVIDIA GTX1080 Ti hardware and Ubuntu18.04 operating system,the detection speed reached 21 FPS to 43 FPS.(3)A convolutional neural network model compression algorithm for adaptive adjustment of pruning position is proposed.The algorithm consists of two parts: channel pruning based on attentional mechanism and model-dependent pruning strategy based on reinforcement learning.Firstly,experiments are carried out on the open data set CIFAR-100,and the pruning algorithm proposed in this paper is used to pry the current mainstream deep convolutional neural networks VGG19 and Res Net56,and their pruning effects are evaluated.Then,the reconstructed model is used to evaluate the generalization of the pruning method on the Image Net data set.Experimental results show that the proposed algorithm can achieve data-dependent adaptive pruning while maintaining high precision of the model,and the algorithm is universal and can be applied to various deep convolutional neural networks.In addition,the proposed adaptive pruning algorithm was applied to the Efficient Net classification network model in the proposed escalator passenger abnormal behavior detection framework.When the total average accuracy and total average accuracy of escalator passenger abnormal behavior classification only decreased by 1.65% and 1.20% respectively,The minimum detection speed and maximum detection speed increased by 38.1% and 44.1%respectively.