Human Action Recognition Based On Depthinformation

Human action recognition is a hot research topic in computer vision,which has attracted the attention of many experts.Depth images are not affected by illumination and other factors.Many methods for human action recognition based on depth information have been proposed successively.How to fully and effectively represent the depth images sequence of an action is the focus of action recognition research.In order to solve this problem,the following two human action recognition algorithms based on the existing methods of action representation are proposed.The main works of this paper are as follows:In order to obtain action information from different perspectives,a human action recognition algorithm based on multi-perspective depth motion map(multi-perspective DMM)is proposed.First of all,the motion history point cloud(MHPC)is generated from the depth image sequence of an action.The rotation matrix is used to rotate MHPC around the Y axis at a certain angle.The distribution of the point cloud become more densely after the MHPC and the rotated MHPC are projected onto the Cartesian coordinate plane.Multi-perspective depth motion maps are computed by the spatial coordinates of the overlapping points.Then,Multiperspective depth motion maps are encoded with the histogram of oriented gradient(HOG).The feature descriptors are into the support vector machine(SVM)for classification.The multi-perspective DMM adds more action information from different perspectives,but the time information during the motion is not fully utilized.To solve this problem,a human action recognition algorithm based on spatio-temporal energy maps is proposed.First of all,MHPC and the rotated MHPC are projected onto the Cartesian coordinate plane,and spatiotemporal energy maps are generated by space coordinate and temporal coordinate of the overlapping points in projected MHPC.Compared with the multi-perspective DMM,spatiotemporal energy maps as three channels color image contain more information during motion.Especially the time energy and space energy are fully captured in the spatio-temporal energy maps.The feature extraction of spatio-temporal energy maps is similar to the multi-perspective DMM,but the classifier has been improved.The SVM has been replaced by the more efficient extreme learn machine based on kernel(KELM).Experiments of the above two algorithms of human action recognition were carried out on the public dataset and the self-built dataset.The results show that both the multi-perspective DMM and the spatio-temporal energy maps improve the accuracy of action recognition effectively.Compared with the algorithm of multi-perspective DMM,the algorithm of spatiotemporal energy maps has better recognition effect,indicating that spatio-temporal energy maps containing both time information and spatial information represent the action samples more comprehensively and efficiently.