Research On Fall Behavior Recognition Algorithm Based On Deep Learning

With the promotion of deep learning technology in various fields and deeper mining in these fields,its effective applications are also increasing.Using intelligent video surveillance technology based on deep learning technology to solve various problems caused by the aging population has become a very valuable research direction.Among them,the images or video sequences captured by various video surveillance devices are valuablely extracted and processed,analyzed and understood,in order to solve the increasingly serious problem of "falling" of the elderly has become a very practical topic.Compared with the current common fall detection system based on various sensors or wearable,it is more convenient and practical,easy to modular large-scale deployment and better compatible with other intelligent functions of video monitoring.Therefore,it is meaningful to recognize abnormal human behavior through video surveillance,especially the recognition of human falling behavior.This paper aims to use different depth learning technology to identify and optimize human fall behavior,mainly from two aspects of research,and research and experiment on these two aspects.The main contents of this paper are as follows:(1)Research on the recognition method of fall behavior of spatio temporal graph convolutional neural network based on 2D skeleton information.First of all,using the current effective pose estimation algorithm to extract the joint points of the human body from the RGB video,and then arrange,optimize and expand the joint points of the human body according to the graph network structure in the time dimension and the space dimension.Introduce spatial attention mechanism,use the representation based on body parts to understand the importance of each part and their relationship in space and time,fuse local information and enhance the characteristics of different action behaviors,through the spatio temporal graph convolutional neural network Learn and train 2D skeleton information,and finally classify and predict the features extracted from the network to improve the accuracy of falling behaviors.(2)Research on the fall behavior recognition method based on RGB image information and efficient displacement convolutional neural network.The feature displacement operation on the time axis allows the neural network to learn interactive spatio temporal features,coupled with an efficient convolutional neural network can be better deployed on the hardware than the 3D convolutional neural network in the case of a small performance gap.Compared with another type of recognition algorithm based on Two-stream convolutional neural network,the recognition accuracy performance can be greatly improved.