Methods For Detecting Human Pose And Recognizing Human Action In Video

Recognition of human actions from video sequence is a highly interesting area of research in computer vision,and more and more researchers are interested in the research area recently, due both to the number of potential applications and its inherent complexity. From an application perspective computer vision-based methods often provide the only non-invasive solution making it very attractive. The research area contains a number of hard and often ill-posed problems such as inferring the pose and motion of a highly articulated and self-occluding non-rigid3D object from images. This complexity makes the research area challenging from a purely academic point of view.As an important research field of computer vision,there are a lot of achievements of human action recognition in recent.Most of these research are confined to the daily gentle action of human,such as walking,jogging, sitting down etc.In sports video,there are many specific characteristics about human actions, such as the upside down pose of human, high speed,moving viewpoint.In this thesis, we focus on action recognition from sports video. In order to recognize human action in sports video,we have done some research works on the following issue:human detection in sports video,Foreground extraction,object tracking and representation and recognition of human action.The main contributions of our work are listed as follows.1. We present an efficient approach to detecting upper body in unconstrained Environments based on Shape Context and Histograms of Orient Gradients. The method contains two steps. Shape context matching is used to get the candidates of upper bodies in the images at first, and then validation of each candidate is performed by using the Histograms of Orient Gradients.The method has two advantages than existing body detection methods. One is that it can detect upper bodies in all kinds of postures other than upright standing posture,the other is higher performance than HOG detection method. 2. We present an approach to extracting human body from an image without interaction of user. Grabcut is an Interactive segmentation algorithm, which is employed to extract the foreground of an image, but it works wrong frequently when the pixels around the foreground are similar to the pixels of the background. We advance a new cost function to correct the mistake.Using the human detecion method,we presents a human body extraction method from image without interaction.3. Mean shift is an effective tracking algorithm, which is employed to track moving object in video, but it loses object frequently when the tracked Object in the video moves rapidly. We advance an approach to tracking object moving rapidly, and it can work effectively when the object is in cluttered environment.4. Using the HOG of foreground image to represent image features,we construct a Self-Similarities Matrix to represent the human actions in different viewpoint.The new SSM are more suit to representing and recognizing the human action in sports video which are shot by a moving camera.To conclude,in order to extract human body from video,we advance a method for human detection,a method for object tracking and a method for human body segmentation.By use of the three methods,human body can be extracted from video,and we can construct a self-similarity matrix based on HOG features of Foreground to represent human action. This presentation of human action is not affected by different viewpoints and camera movement.