Human Action Recognition Based On Kinect Skeleton Information

In the field of machine vision, the research on Human Motion is studied all the time, from human body detection and tracking to human posture recognition and action recognition, and even behavior understanding. At the beginning, much research work on common2D images including Gray and RGB images has been done, and various kinds of advanced image processing algorithms have been designed, but there are also some noise problems that cannot be avoided, like illumination, color, texture, and overlap, etc. With the development of hardware and academic theory, researchers are not confined to the common2D images, and seek for new presentations of image, like X-ray and infrared images, and hope to find data which common images don’t have. In recent years, Depth maps come into the public notice, and its pixel is the distance between camera and a point in the space. In other words,3D information can be directly obtained from3D space, and it will be convenient for researchers in comparison to3D reconstruction from multiple images. In fact, depth maps also belong to computer vision, but the difference lies in its data presentation, so many current image processing algorithms can be used and improved. In this thesis, human action feature model and recognition algorithm are studied based on depth maps, and the research achievement will provide a new method for human action recognition. The main contributions are as follows:(1) A summarization of research background, significance and status is introduced. The problem research work faced and the feasibility of depth maps in solving the problems are analyzed, which provides a new ideal for this thesis.(2) The Microsoft Kinect camera is used as acquisition device to obtain depth maps, and get20human body joints with the help of released SDK. On the base of them, the kinetic feature of joints is analyzed, and then a joint-angle variation series as the feature model is proposed. The model is not complex, easy computing and composed.(3) The main recognition algorithms are described. As to designing the classifier, the Dynamic Time Warping algorithms is used after analyzing the characters of feature model indicated above, instead of just choosing a complicated and advanced algorithm simply. However, the traditional DTW algorithm does not provide the ideal result because of its drawbacks, so a improved method is proposed in order to enhance the matching accuracy. (4) Considering the time cost of template matching, parallel computing is adopted to accelerate the speed. A large number of experiments are performed, and the results show the proposed method for human action recognition is effective and robustFinally, a conclusion is made for whole contents of this thesis, together with the perspective of this filed for the next step.