Unusual Behavior Detection from Video Images Using Histogram of Oriented Gradients of MHI and k-Nearest Neighbors

Abstract

This paper proposes a method of detecting unusual human behavior from video images for automated visual surveillance. The method detects unusual human behavior by learning examples of usual behavior and then detecting behavior that is different from the usual. Histogram of oriented gradients of Motion History Image (MHI) is used for describing the features of human movements, and k nearest neighbors is used for the classifier. The performance of the method was evaluated by applying it to unusual pedestrian behavior detection on a street. As a result, true positive rate was 85% when false positive rate was 3%.