Research On Human Action Recognition In Videos Based On Deep Learning

Human action recognition is an important research in the field of artificial intelligence and computer vision.In our daily life and work,intelligent devices with human action recognition capabilities have important applications in human-computer interaction,video surveillance,unmanned driving and other aspects.At the same time,there is an increasing demand for such products.Therefore,the research of the human action recognition in videos has crucial value and significance.This thesis mainly focuses on human action recognition in videos,involving key technologies of using deep learning to achieve it.The specific work is summarized as follows:1.To solve the problem of action recognition on the limited set of human action video samples,a more reasonable and pertinent two-stream convolutional neural network is designed at first.Then,in order to increase the convergence speed and reduce the degree of over-fitting during network training,the Maxout activation function is introduced to replace the ReLU function to perfect the initial network.Finally,in order to further improve the recognition and generalization ability of the network,the Stochastic Pooling method is used to perfect the pooling operation of the network.The experimental results show that the two-stream convolutional neural network designed in this thesis can effectively accomplish the task of recognizing human action in videos.2.Aiming at how to make full use of the video action features extracted by the two-stream convolutional neural network to improve the recognition ability of the network,a scoring feature fusion method based on the two-stream convolutional neural network is proposed.Firstly,the types of scoring features and the timing of fusion in two-stream convolutional neural network are analyzed.Secondly,two specific fusion methods are designed for different stages of feature fusion in the network,that is,multi-feature fusion method after single-frame spatio-temporal feature extraction and the two-stream feature fusion method after video single-stream feature extraction.Finally,the two-stream convolutional neural network designed in this thesis is used in KTH human action datasets to accomplish the task of recognizing human action in videos by combining the above two feature fusion methods.The experimental results show that the scoring feature fusion method designed in this thesis can effectively increase the recognition rate of the two-stream convolutional neural network.Among them,the network recognition rate obtained by using the two-stream feature fusion method after video single-stream feature extraction is optimal,and the recognition rate is better than other traditional algorithms.