Research On Human Action Recognition Based On Depth Maps

Human motion recognition has been a hotspot in the field of computer vision,and has attracted wide attention from academia and business circles.The combination of depth camera and machine learning reduces the difficulty of target detection and segmentation.They also provide new ideas for pose estimation and human motion recognition.How to use depth data to improve the performance of human motion recognition system is an important problem to be solved urgently at present.In this thesis,the subject of human motion recognition based on depth image is studied.The research contents include low level feature extraction,mid-level feature coding,motion classification model and so on.The main contributions of this thesis include:(1)A human motion recognition method based on the trajectory of skeleton joint points is proposed.In order to improve the accuracy and real-time performance of motion recognition,a motion recognition method based on joint trajectory is proposed.Inspired by the human motion experiment in psychophysics,the human motion is represented by the trajectory of the joint points of the human skeleton,which can fully express the human motion in the space-time dimension.On this basis,the motion trajectories of joint points are clustered by using the Gauss mixture model,and then the feature quantization is carried out by Fisher vector.Considering the real-time requirement of action recognition task,an action recognition method based on extreme learning learning machine is proposed to improve the real-time and accuracy of action recognition task.Finally,experiments on public datasets demonstrate the effectiveness of the proposed method.(2)A human motion recognition method based on the sequence of inter-articular angles is proposed.Aiming at the problem of complex background and viewpoint change in human motion recognition,an action recognition method based on the sequence of joint angle change is proposed.Inspired by mechanism and robotics,human motion is represented by the sequence of angle changes between adjacent joints and non-adjacent joints,and then classified by k-nearest neighbor classifier.Because the action duration of different individuals is different,dynamic time warping algorithm is used to calculate the distance between samples.Finally,experiments are carried out on UTD-MHAD and KARD datasets,and experimental results show that the method is effective.(3)A human motion recognition method based on depth motion projection and temporal domain segmentation is proposed.This thesis presents an effective method for human motion recognition based on depth motion maps(DMM).In order to overcome the shortcomings of traditional feature descriptors,Gabor filter banks are introduced to extract the features of DMM.In order to make up for the shortcomings of DMM in temporal domain expression,a strategy of temporal domain segmentation is proposed to describe the temporal domain information.Finally,classifiers are used to classify human movements.The experimental results verify the effectiveness of this method.(4)A human motion recognition method based on 3D motion history image and multi-task learning is proposed.Aiming at depth image sequence,a human motion recognition method based on Gabor feature extraction and multi-task learning is proposed on the basis of 3D motion history image.In order to solve the problem of inadequate representation of motion history images based on contour features,Gabor filter banks are introduced to extract features from 3D motion history images.In order to depict the changing process of human motion in different temporal dimensions,the strategy of temporal domain pyramid is introduced to partition action video.Finally,in order to mine the correlation between action recognition tasks,a multi-task learning training action classification model is introduced.The experimental results show that this method is obviously superior to the existing methods.The research results of this thesis can provide theoretical and technical support for further research and application of human motion recognition.