Multi-layer Perceptron Based Video Surveillance System

Abstract

Conventional video surveillance systems involve continuous human monitoring of video feeds which is laborious and error prone. In this paper, we present a design for automation of video surveillance using behavior analysis and face recognition. Behavior analysis involves extraction of `motion vectors' using multi-object tracking, which are then fed to a multi-layer perceptron. This is shown to be more adaptive than conventional techniques, requires lesser training time and gives improved performance as a result of requiring simpler networks as compared to typical deep network based approaches. Experiments conducted with two real world scenarios gave accuracy levels of above 75% and performance improvement of 7% against contemporary solutions. Face recognition involves facial feature vector extraction and uninformed search in the feature space guided by a Kohonen classifier, eliminating the need for deep-networks while requiring lesser computation than legacy techniques. Face recognition has been tested against the `Landmarked Faces in the Wild' dataset.