Adaptive Silhouette Extraction and Human Tracking in Dynamic Environments

Extracting a human silhouette from a video sequence is the enabling step for many high-level vision processing tasks, such as people tracking, video surveillance, and mobility assessment. Silhouette extraction schemes are widely explored in the literature, but most approaches work efficiently only in constrained environments where the background is relatively static. How to accurately and effectively model and update the background and how to deal with shadow and objects attached to human body are still two significant challenges for silhouette extraction. In this paper, we address these challenges in real-world unconstrained environments where the background is complex and dynamic. In the algorithm proposed, we extract features in a color space, accumulate the feature information over a short time, fuse high-level knowledge and low-level feature information, and then build a time-varying background model. A fuzzy logic inference system is also developed to detach silhouettes of moving objects from a human body silhouette. Our extensive experimental results show that the algorithm works efficiently and robustly.