Research On Deep Learning Based Human Action Recognition

Human action recognition which belongs to computer vision is an important issue,but also a great research foundation of artificial intelligence.It has been widely used in monitoring security,entertainment facilities,smart appliances and so on.The subject is committed to the computer to capture the video sequence,so that the computer could recognizes human action.However,the occlution,shadows,complex background,jitter camera,changing of light,different distances between the characters for the camera which causes the texture variation,scale variation,the diversity of action style and so on is the main problems while facing this study.Nowadays,deep learning has been extending to so many areas like image processing,speech processing,translation,recommendation,evaluation,prediction analysis and so on.However,in the analyzing of continuous video sequence field,how to apply deep learning to human motion recognition is a potential research topic.All the research in this paper stand in this perspective that human action recognition tasks for different input format are based on deep learning in different ways.Convolution Neural Network act as a deep learning method has been widely used in image recognition and classification.Its biggest advantage is not only the high recognition accuracy,but also the ability for not depending on the traditional hand-craft features,which could automatically learning features from raw data.In this paper,focusing on different video input format,this paper proposed different action recognition methods and frameworks.a)To deal with skeleton squences,two skeleton mapping methods for skeleton sequences are proposed: skeleton trajectories maps and skeleton distribution maps,in which the convolutional neural network is adopted to automatically learning the maps and do the classification task.The skeleton based methods are evaluated on G3 D,UTD-MHAD and MSRC-12 datasets and the experiments results show that the proposed method acquired the state of the arts accuracy.b)Focusing on depth data,this paper firstly adopt Histogram of Oriented Displacements(HOD)and Dirichlet Process Mixtural Model(DPMM)to segment the action sequences and build the action dictionary.Then,using CNN to classify Depth Motion Maps(DMM),which ensures that actions are coding by the dictionary reasonable.Finally,an HMM-SVM classifier is adopted to recognize the action.The proposed framework acquire 100% accuracy on MSRAction-Pairs dataset and also performs well on MSRDaily Activity3 D dataset.c)As for traditional RGB video sequences,we not only proposed a novel action recognition method,in which optical flow throw the whole video sequence are encoded according to the temporal character and then mapping to one figure for CNN to learn and classify,but also we extended this method to a real-time Unmanned Aerial Vehicle(UAV)application.We proposed a robust gesture recognition based Robot Operation System(ROS)which includes 5 gesture commands,and the average recognize accuracy is more than 93% with the distance range from 5m to 60 m.What's more,the real-time system can feeds back the pilot within 0.4s.The recognize accuracy,recognize distance and the recognize speed are advanced in the word.