Ma-Th algorithm for people count in a dense crowd and their behaviour classification

Abstract

In this paper, an intelligent surveillance algorithm for estimating the people count in a crowd and also classifying the crowd behavior as normal or abnormal is proposed. This method combines the machine learning and threshold based algorithms (Ma-Th) to estimate the people count and crowd behavior analysis. First, the foreground is segmented using ViBe algorithm. Subsequently, the features are extracted using bounding box characteristics such as crowd density, relative height/width, foreground pixel's horizontal/vertical mean. In addition to that the foreground pixel's kinetic energy and crowd distribution are thresholded. These features are learnt by Relevance Vector Machine (RVM) learning algorithm for both people count and their behavior classification. Experimental results obtained by using benchmark surveillance datasets such as Pets 2009, UMN, UCSD and videos downloaded from internet show the effectiveness of the proposed algorithm.