Human Action Recognition Based On Delaunay Skeleton Model

Human action recognition based on vision is one of the most promising applications of computer vision.In recent years, this problem has caught the attention of researchers from industry, academia as well. Human action recognition is the extraction of concise low-level features form input video or sequence of images;and then describe midlevel action from low-level features;in the last interprete high-level semantic from primitive actions.This paper presents a human action recognition system based on delaunay skeleton model for single person's action recognition in simple background. First we obtain the human's moving foreground by background subtraction method. And then we use delaunay trangulation to build delaunay skeleton model.and add energy parameters in the model to improve the recognition rate effectively and achieve the purpose of multi-feature recognition. In the last we use the hidden markov model to achieve a complete human action recognition system.First, we detect the moving foreground and convert it to binary image. Through analyse several common moving foreground detection algorithms' principles. advantages and disadvantages, we use the background subtraction method to detect moving foreground, and use morphology method to the binary images' post-processing.Second we build the delaunay skeleton model. We obtain the human's contour by morphology method. The contour extraction algorithm in morphology method has the advantage of simple and effective, and it's time complexity and space complexity is low.We obtain the centroid of human body based on the coordinates of all points on the human body's boundary. We obtain the maximum point set and minimum point set of the function by calculate the distance between all the points of boundary and centroid. We obtain the discrete points on the boundary by the number of points between each two extremum points. The discrete points is essential to the delaunay skeleton model. We trangulate the human foreground by the classical quickhull algorithm, remove the triangles outside the contour and save ohter triangles. Through a traversal algorithm written by myself, we traverse all the saved triangles to establish the delaunay skeleton model and obtain the triangle's center of human's torso, hand, foot. We obtain the energy parameter by the the triangle's center of human's torso, hand, foot and the standard parameter of human body part, and add the energy parameter in the skeleton model to compise the final feature matrix. The feature matrix is matched with the symbol table and convert to symbol finally.Finally we introduce the basic principle of hidden markov model, and it's key algorithms.Hidden markov model is a wide application in human action recognition field. It is simple and practical for programmers. Finally we show the experiment results of the applicantion of hidden markov model.