Classification and quantification of occlusion using hidden markov model

Abstract

Over the last few years, gait recognition has become an active area of research. However, one of the shortcomings is lack of a method for quantifying occlusion in scenes used for capturing gait of individuals. Occlusion can occur primarily because of two reasons. Firstly, movement of certain body parts of a human being occludes some other body parts of the same human, which is called self occlusion and secondly, occlusion of the body parts caused by some other human being. The objective of this paper is to quantify occlusion of different parts of the human body using Hidden Markov Model (HMM) and classify the scene of occlusion as one of the three cases of occlusion, namely, self occlusion (single individual moving), occlusion in a crowd moving in same direction and occlusion due to movement of human beings approaching from opposite direction. We train one HMM for each body part relevant for gait recognition. An HMM is a statistical representation of probability distribution of a large number of possible sequences and in the current context these are the sequences of frames extracted at regular interval from a given video. The steps involved in achieving the objective are feature extraction, HMM training and finally the classification or hidden state generation. © 2011 Springer-Verlag Berlin Heidelberg.