Face Occlusion Detection for Automated Teller Machine Surveillance

Real time automatic alarm systems play an essential role in security management, as evidenced by the surveillance cameras installed in nearly all automated teller machines (ATMs). Whereas manual video surveillance requires constant staff monitoring, fatigue or distraction is a common human error. Therefore, this work presents an effective detection system for facial occlusion to assist security personnel in surveillance by providing both valuable information for further video indexing applications and important clues for investigating a crime. A series of methods that include identifying and segmenting moving objects is formed. The moving edge is then captured using change detection of the inter-frame difference and the Sobel operator. Next, a Straight Line Fitting (MSLF) algorithm is developed to merge the splitting blobs. Additionally, a mechanism involving moving forward or backward justification is used to determine whether an individual is approaching a camera. Moreover, the lower boundary of a head is computed, followed by use of an elliptical head tracker to match the head region. Finally, skin area ratio is calculated to determine whether the face is occluded or not. The proposed detection system can achieve 100% and 96.15% accuracy for non-occlusive and occlusive detection, respectively, at a speed of up to 20 frames per second.