Research On The Key Technique Of Simple Action Recognition In Video Sequence

In the field of computer vision, action recognition is a hot topic and has being widely explored recently. It is a valuable theoretical problem and a high technique, which can also be applied to many domains, such as intelligent surveillance, robot, human-computer interface etc. This paper mainly focuses on the research of posture modeling and simple action recognition based on spatio-temporal interesting points (STIPs). Simple action can be some common actions of in our daily life, for example: walking, running, waving, skipping, and bending.Posture modeling is a critical step for action description and recognition. The contour of the people can not be extracted completely in the process of detecting and tracking, in order to avoid these disadvantages, a new posture modeling and action recognition method is proposed in this paper. First of all, STIPs are extracted from the learning samples, and they are used to describe the human moving character. In fact, one posture consists of a set of STIPs; Secondly, a unsupervised clustering method NERF C-Mean is adopted to classify salient postures from these posture samples; Thirdly, a GMM model is established for each posture clustering result; so the posture information of the moving people can be received, and the process of moving human detecting and tracking can be substituted.Action recognition is on the high-level semantic understanding. Simple action recognition process is based on the appearance order of salient postures in the test video sequence, finding the max probability path in salient posture transition graphic with compution, then identifying the action classification which it belongs to. So the premise step is that transitional probability among salient postures should be calculated first, and a Visible state Markov Model(VMM)is learnt to describe various human actions. The Bi-gram method is put forward for action recognition, which is primarily used in Chinese characters classification, and it can better understand the high-level semantic meaning. Extensive experiments are conducted and the results prove its robustness and validity for different background and different scaled videos and for occlusion problems as well.