The Study Of Human Action Recognition Based On Kinect Depth Image

Human action recognition has a wide application in video surveillance, human computer interaction, and gaming and entertainment industry.The traditional two-dimensional image recognition method is susceptible to interference complex background, lighting changes, shadows and objects and other factors,whose practical application suffer a lot restricted.In recent years, with the advent of the low-cost depth of the camera, represented by Microsoft’s Kinect, setting off a wave of international research on pattern recognition based on the depth of the image.The depth map directly reflect the surface features of the three-dimensional image so that overcome the difficulties encountered by the conventional image motion recognition technology.Therefore, the integration of two-dimensional image and depth image feature representation and classification method for improving the usefulness of action recognition has high research value.Using Kinect depth image-based human action recognition technology we have implemented an algorithm to describe human action, and we get a new descriptor of human action. First, we use the projectors to initially quantify the four-dimensional space, then extract the distribution feature of body surface normal vector on the projectors from the depth sequence to construct a histogram, whitch is regarded as the descriptor of human action.In order to further improve the recognition rate, we learn the initial projectors’density from which we figure out the high discriminative projectors with the help of the weight of support vectors in the SVM model.We generate a new set of projectors by adding disturbance factor to the high discriminative projectors.Last we re-construct the descriptor of human action based on those new projectors.We use the generic international human motion data set that are MSR Action3D set and MSR Gesture3D set to test our algorithms by cross-validation. Experiment shows that, our algorithms outperform2%on accuracy compared to HOG3D proposed by Klaser at el and DMM-HOG proposed by Yang at el.In the MSR Gesture3D set, our algorithms outperform4%on accuracy compared to HON4D proposed by Omar at el.This shows the feasibility and robustness of our algorithm.