Face pose analysis from mpeg compressed video for surveillance applications

Here, we propose a hierarchical approach for face pose analysis, which works with MPEG compressed video of head-and-shoulders style. The proposed approach consists of two layers: coarse human body tracker & face locator and fine face tracker. To robustly estimate the parameters of face pose, we use transform domain features to detect and track the reliable parts of the human body for removing distractive factors (e.g. skin-toned nonface objects, random cluttered caused by moving objects). Within the tracked body part, we apply the high-precision face tracker to estimate the pose parameters. The proposed approach is found to be useful particularly for automated transaction service (ATS) surveillance, where clear face is an important visual cue when searching and browsing a video database. The frontal face is practically considered a clear one. A user-study is conducted to show the effectiveness of clear faces selection on automated teller machine (ATM) testing clips.