A novel efficient method for abnormal face detection in ATM

Abstract

Face detection in video is a challenging and interesting topic, especially when it's applied in Automatic Teller Machine (ATM). We propose an efficient algorithm to detect arbitrary occluded faces in ATM surveillance. A novel and time-saving foreground extraction method is proposed to obtain accurate foreground. After that, we locate the face with two cascaded steps. An empirical rule-based face localization is utilized to locate the face roughly, then adaptive ellipse fitting helps accurately locate the face. In order to detect the occluded faces, we use ADABOOST to combine a skin color detection and face templates matching. Experimental results show that our algorithm achieve a high detection rate while keep the false negative rate pretty low.