Research On Human Action Recognition Based On Vision

Research on human action recognition based on vision is an important researching content of computer vision. The technology offers broad applications in fields of the intelligent monitoring, human-computer interaction and virtual reality, motion analysis, video annotation and so on. However, there are still some existing problems in this research field, such as human occlusion and self-occlusion, as well as action ambiguity caused by projecting 3D space to 2D image plane. That is why technology of human action recognition has been quite a challenging research topic. Based on this, this paper concentrates on human action recognition based on vision, focusing on human motion detection and motion analysis based on single-camera vision and human action recognition based on RGBD vision. In addition, this paper puts forward some solving methods and improvement measures. The main innovations of this paper are as follow:(1) In the single-camera vision, this paper proposes an improved method for human motion detection and motion analysis, which includes two main parts as object motion detection and motion analysis. In the object motion detection part, this paper proposed an improved detection algorithm, which first changes the updating formulas of the original Gaussian mixture model, combined with three frames subtraction to extract 2D contour image of human body. In addition, this paper applies this algorithm to common behavior databases for testing. In the motion analysis part, by computing the aspect ratio of human contour, the area of the rectangular box and the offset of centroid as human action feature information, the judgment basis of action types is in accordance with time-varying variables of these information.(2) In the RGBD vision, this paper proposes two action recognition methods based on no-training and training model respectively. In the human action recognition method based on no-training model, the characteristic of pointing movement time for skeleton joints is used to recognize different types of human action in real time. In the human action recognition based on training model, this paper first selects the relative positions of skeleton joints and joint angles as human action features, establishes an action database containing different kinds of human action, then trains and classify these feature data by utilizing multi-class support vector machine classifier, finally recognizes human action type. Experimental results show that these two methods for human action recognition can accomplish recognition task.