Pictorial structure based people detection and pose estimation in videos

An open vision problem is to detect the people and automatically estimate the body configuration. Finding people and localizing their body parts is hard because people can appear in variety of clothes, pose and they can move unpredictable. Numerous solution and methods have been developed over the years and often this address some special cases. One proposed direction to solve the non-rigid object detection problem is the Pictorial Structure based methods. In this paper, we present a generic framework for people detection and pose estimation based on the Pictorial Structure. In our framework we combine the pictorial structure with a strong part detector, and we use time constrains to speed up and enhance the entire detection process. We demonstrate that using the time constraint does not reduce the generality of the proposed framework, but in most cases this constraint seed up the entire process. Even in the worst cases when between consecutive frames is no connection, the framework is working correctly.