Optical Flow-Initiated Particle Filter Framework for Human-Tracking and Body-Component Detection

The principal task in any tracking system is localizing the target motion to minimal bounds, covering target region inside successive video frames. The accuracy of any such system is significantly affected by the initialization parameters. For proper initiation and tracking outcome various methods for target-modeling have been proposed over the years. Our present work focuses on estimation of different body-component positions in a monocular video sequence. We propose a robust particle filtering framework based on optical flow for tracking human in a cluttered scene. The combination of optical flow with the particle filter framework obviates the need for any prior information on target characteristics as the initiation-information is generated online for individual targets. Furthermore, optical flow method provides good initial estimates of target-size and motion characteristics for the prediction stage in particle filter framework. Afterwards, bottom-up search for body-templates has been performed to identify the locations of various body components.