Research on Recognition Algorithm of Abnormal Behavior of Workers in Two-Stream Convolutional Network

Abstract

In order to identify the abnormal behaviors of workers in the factory, this paper proposes an improved algorithm for identifying abnormal behaviors of workers in a two-stream convolutional network. The workers’ body contour shape information extracted from the convolutional neural network is input into the LSTM network in order to extract timing information between frames. Secondary extraction of the dense optical flow image of the video image, sparse extraction of pixels with small changes in optical flow value in the dense optical flow image, and then put the new continuous optical flow image into the continuous optical flow image network. The two networks are fused after softmax classification to get the final recognition result. Experiments on the CAVIAR dataset, CASIA dataset, and self-built behavior dataset show that compared with other abnormal behavior detection methods and traditional two-stream convolution algorithms, the accuracy of the improved algorithm in this paper is improved by 1% -4%.