Intelligent visual surveillance, which relies heavily on human motion detection / recognition and people recognition, has received a lot of attention for its use in effective monitoring of public places. However, there is a concern of loss of privacy due to distinguishable facial and body information. To deal with this issue, researchers proposed to protect privacy example by filtering of face or body areas, and developed methods of people identification from videos in which people's faces has been obfuscated, masked by digital filters. Along the same line of research dealing with videos in which the people faces were masked by filters, this paper introduces a method to detect abnormal behavior. In the proposed method, we first mask face areas in videos by Multiple Instance Learning tracking, and extract silhouette area from each image. We then extract features using affine moment invariants, and perform classification. We build a database including normal and abnormal behaviors, and we show the effectiveness of the proposed method on cases from the database.