Human Action Recognition Research Based On Kinect Skeleton Data

With the development of modern technology,people's requirements of intelligent life are increasing at the same time.Computer vision and pattern recognition as an important technology to promote human life intelligent,has gained wide attention and rapid development.As an important research direction of computer vision,human action recognition technology involves many fields,has huge use value and application prospect,a lot of researches have been conducted on this issue,and the issue has became one of the most popular topic in this field.Treatment of early action video are RGB color image video based,the color image are two-dimensional data,the image itself vulnerable to light effects,therefore,to deal with it not only troublesome and the effect is not good,so,people agree to find a new way of presenting image.With the appearance of Kinect we can get depth image,depth image includes depth information and skeleton information.Based on the Kinect skeleton information,this thesis have done some research about human action recognition,the main contents include:First,state the research background and meaning of human action recognition,analysis of the status at home and abroad in the field of human action recognition,and find problems remain to be solved.Grasp the specific process of action recognition,in-depth analysis of object detection,feature representation,feature extraction,classification and recognition.Understanding and research the current human motion data sets and work principle of Kinect.Second,this thesis firstly puts forward an algorithm of human action recognition based on local spatial information of skeleton.This method firstly use skeleton data extracted spatial location information of human skeletal joints,space angle and joint angle information,then putting into hidden Markov model for classification and identification.Through experiment on the MSR3 D Action,the recognition results proves that this method can get better result for the complex action sequencerecognition,and in the cross test difficult relative to existing methods has some improvement.Third,taking the timing of action sequences into account and proposed a human action recognition algorithm of bag of features model of skeleton based.this method based on extracting of local spatial feature,adds feature between the frames of image sequences,and use bag of feature model to coding action feature and representation,for better use the temporal characteristic,this algorithm introduces temporal pyramid matching model to further divide action feature to get the last action descriptor,then using the support vector machine for classification at the last.Experiments on current data set shows better recognition result.