Model-based recognition of human walking in dynamic scenes

In numerous content-based video applications, it is important to extract from a video sequence a representation for humans in motion. For example, in generative video (GV), one needs to construct accurate world images for moving objects. Because humans are not rigid objects, this task is difficult. We propose here a model-based recognition of human walking in dynamic scenes. We model the human body as an articulated object connected by joints and rigid parts, and the human walking as a periodic motion. We determine the posture by using a recognition algorithm that estimates the period and phase of walking. We obtain promising results when testing our algorithm with real video.