Deep Learning Based Human Action Recognition

Human action recognition has always been a very important and active research field in computer vision.It has broad application prospects in the fields of video retrieval,intelligent security and human-computer interaction.In recent years,with the decreasing cost of data storage,continuous improvement of computer performance and rapid development of GPU industry,two bottlenecks of deep learning(a large number of training data;huge network parameters and computational complexity)have been well solved.Thus,the research of human action recognition based on deep learning has developed rapidly and become the mainstream.This paper mainly studies human action recognition algorithm based on deep learning.The specific research contents are as follows:Firstly,the dissertation discusses the research background and significance of human action recognition algorithm based on deep learning.The dissertation reviews the development history and research status of human action recognition based on RGB data and color video respectively.The dissertation also discusses the shortcomings of current human action recognition algorithm based on deep learning.In view of these shortcomings,the research content of this paper is proposed and introduced.Then,we introduce the basic knowledge of deep learning,including the basic structure of neural network and the principle of gradient back propagation algorithm.We enumerate some commonly used loss functions and elaborate convolutional neural networks and long-term and short-term memory networks which are often used in human action recognition.Next,in view of the fact that the spatiotemporal features of human action in complex and diverse environments are susceptible to interference and there are a large number of redundant frames in video.Dissertation proposes a human action recognition algorithm based on trajectory weighted depth convolution rank pooling descriptor.This descriptor is computed by convolutional neural network(CNN),trajectory attention map,cluster pooling and rank pooling methods.It can effectively describe the human action dynamics in long-term redundant video in complex background.Support Vector Machine(SVM)is used to perform human action recognition experiments on SDUFall dataset,Weizmann dataset,and HMDB51 dataset.Good results have been achieved on all datasets.Finally,the dissertation proposes a human action recognition algorithm based on stacked unjitterring dynamic image and 3D convolution neural network to solve the problem of the movement of the camera in video,which will cause the quality of dynamic image and the parameter increment of 3D convolution network.By estimating camera jitter in video when encoding dynamic images,the quality of dynamic images can be improved.At the same time,the unjitterring dynamic image is stacked and input into the 3D convolution neural network,so that the 3D convolution network can obtain longer video input without increasing the parameters.