The Research On Detection Algorithm Of Abnormal Behavior Of Industrial Workers

With the continuous development of industrial technology,industrial production is gradually modernized and intelligent,but it still needs to be completed manually in many production links.At this time,workers will be in a dangerous state,and accidents may occur.Therefore,workers' Abnormal behavior is a hot research issue in modern industrial production.The emergence of surveillance video technology has been realized.When workers are in accidents,ambulance personnel can timely find out and make countermeasures.Therefore,it is of great practical significance to realize timely and accurate detection of abnormal behaviors of workers.However,the existing surveillance video technology has two problems.On the one hand,it can only store the accident process of workers and cannot detect or predict the abnormal behavior of workers in real time.On the other hand,the current surveillance video still relies on manual supervision and manual identification of labor.The intensity is subjective.Therefore,the research on the detection of abnormal behavior of workers has received widespread attention.Based on the dual-stream convolutional network,this paper analyzes the two problems in the original dual-stream network where the video image frame network lacks time-series still video image correlation information and the continuous optical flow image network makes it difficult to distinguish similar behaviors,and improves the dual-stream convolution.Product network algorithm to study the abnormal movement detection of workers in the factory.The innovations of specific work and articles are as follows:1.This paper introduces the combination of LSTM network and video image frame network in traditional dual-stream convolution to realize the extraction of time series information in video images.According to the characteristics of the LSTM neural network that can process data that is highly related to time series,the body contour shape feature information of a representative worker extracted from the convolutional neural network is input into the LSTM network to extract frames between frames.Time sequence information,to enhance the understanding of workers' behaviors in video images.2.In order to distinguish the continuous optical flow network from similar actions in the dual-stream convolutional network,before performing continuous optical flow image network training,perform a secondary extraction of the dense optical flow image of the video image to convert the optical flow Pixels with small changes in value are sparsely extracted,and a new continuous optical flow image is input to the continuous optical flow image network for training,so as to obtain the worker's motion information.3.Test the improved algorithm of this paper on the CASIA dataset,CAVIAR dataset,and self-built dataset.The improved algorithm improves the accuracy of workers' prone to abnormal behavior detection based on the original network.It not only improves the recognition rate of abnormal behavior of people in the self-built data set,but also shows a more accurate recognition effect in comparison with similar algorithms.Compared with similar algorithms which must be run in GPU environment,this method is more practical.