Research On Human Action Recognition Algorithm Based On Depth And Skeleton Information

Human action recognition is one of the important issues in the field of computer vision.It plays a significant role in the fields of intelligent video surveillance,intelligent medical care,motion analysis,intelligent human-computer interaction,virtual reality and more.In past decade,though human action recognition technology has a significant development,it has external and internal interference factors including target posture change,shade,lighting,camera dithering,individual behavior difference and etc.Designing a robust human action recognition algorithm is still a challenging task.In this paper,complexity of human action recognition are discussed.On the basis of the latest research achievements in the field of human action recognition,further research based on human action recognition algorithm which is based on the depth and skeleton information.The main research work is as follows:(1)In order to overcome the drawbacks of human action recognition algorithm based on depth data,this paper improves the algorithm from three aspects.Firstly,the traditional feature extraction methods based on depth information don't have spatio-temporal restraint and are easy to generate redundant data.Therefore,to effectively extract spatio-temporal interest area,one effective extraction method of the region of area(ROI)is put forward.Secondly,in order to solve the problem that traditional feature extraction methods are likely to ignore the frame's motion information,the method of histogram of spatio-temporal oriented principal components(HSTOPC)is proposed.Thirdly,in order to solve the problem that the traditional temporal segmentation models are likely to lose the information of the partition,multilayer time domain segmentation model is proposed.The experimental results of the standard database show that the algorithm has good robustness.(2)In order to overcome the drawbacks of human action recognition algorithm based on skeleton data,this paper improves the algorithm from two aspects.Firstly,simple interpolation method can't accurately describe the position of the skeleton joint in the interaction.Therefore,classification interpolation method is proposed to fit the position of frame interpolation accurately.Secondly,traditional feature descriptors can't accurately describe the joint information of skeletal posture.Therefore,quaternion feature descriptor which encodes skeleton information is proposed to represent the skeletal spatial location information.The experimental results of the standard database show that the quaternion feature representation algorithm has good robustness.(3)In order to handle the problem which is the lack of shape information of skeleton data and depth data,this paper adopts feature fusion with deep information and skeletal information.Meanwhile dynamic time warping(DTW)and fourier temporal pyramid(FTP)are adopted to solve the problem of temporal misalignment and noise.The experimental results of the standard database show that the feature fusion algorithm has good robustness.