RVM-Based Human Action Classification in Crowd through Projection and Star Skeletonization

Abstract

Detection of abnormal human actions in the crowd has become a critical problem in video surveillance applications like terrorist attacks. This paper proposes a real-time video surveillance system which is capable of classifying normal and abnormal actions of individuals in a crowd. The abnormal actions of human such as running, jumping, waving hand, bending, walking and fighting with each other in a crowded environment are considered. In this paper, Relevance Vector Machine (RVM) is used to classify the abnormal actions of an individual in the crowd based on the results obtained from projection and skeletonization methods. Experimental results on benchmark datasets demonstrate that the proposed system is robust and efficient. A comparative study of classification accuracy between Relevance Vector Machine and Support Vector Machine (SVM) classification is also presented.