Human Action Recognition Based On Spatial-temporal Features

Human action recognition is a very important area of computer vision, as well as a challenging problem. Recognizing human actions from video sequences is an emerging topic which related to many kinds of research fields, such as image processing, machine vision, pattern recognition, biological scinece and so on. The application of this technology is also extensive, for example, human computer interaction, surveillance, virtual reality and sports analysis. Recognizing human actions from video sequences is a hot topic as well as a difficult research in computer vision these years. Although many scholars in many countries have mad thorough research on this problem, it still has a long way to go for practical applications. Most of human action recognition algorithms at present need large amont of data for training, which bring low real-time performance. To solve this problem, a fast and effective method which applied in simple scene based on key poses has been proposed in this paper.Firstly, some information about foreground segementation has been introduced. After comparing some methods of moving object detection with each other, background subtraction has been choosen for extracting human binary image from background image. We use background subtraction in RGB color space for feature extraction. Comparing with traditional background subtraction using gray images, our method achieves better effects.Secondly, we introduce something abour feature extraction techniques. Then, we choose contour points of human silhouette as feature for human action recognition, and describe the method how we get that. After that, we obtain the trajectories of centre point of each frame, and create a spatio-temporal feature value to describe the motion direction and speed of each action.Finally, the key poses are learned by means of K-means clustering based on the Euclidean distance between each contour point and the centre point of the human silhouette. Combining with the spatio-temporal feature value, here we use a method of ergodic matching that base on maximum similarity for recognition. The proposed method can improve accuracy by distinguishing amphibious poses and increase suitability for real-time applications by reducing the computational cost.We transplanted the algorithm to the MFC platform and using the public available Weizmann datasets with newly created datasets for experiments. The result shows that our algorithm can accurately identify human action from video sequences in a short time...