Anomalies Identification in Images from Security Video Cameras Using Mask R-CNN

Abstract

In this work we developed a system to identify anomalies in images from video security cameras in an urban environment. Initially people are detected in the images using Mask R-CNN. From the binary mask are extracted characteristics of the people so that the anomalies can be detected. In order to facial recognition we used Facial Landmarks so that the system knows the residents and authorized people avoiding the false anomalies. We considered four anomalies in this work: the act of jumping a wall, standing for a long time in front of the residence, walking thru the sidewalk several times and entering a place without permission.