A Large Scale Crowd Density Classification Using Spatio-Temporal Local Binary Pattern

Abstract

Increasing world wide population is leading to dense crowd gathering at public places. Due to mass gathering at large scale, crowd related disaster has been frequently occurred. In order to prevent crowd calamities, automated crowd scene analysis has been a topic of great interest. Density is the status of crowd which is essential to classify in visual surveillance system primarily for security aspects. Most of the existing techniques work on detection and tracking of individuals. Due to fewer pixels per target, multiple occlusion and perspective effects etc., detection and tracking of individuals is a complex task in dense crowd scenarios. This paper presents a novel strategy for large scale crowd density classification powered by dynamic texture analysis. This approach consists of an interest points detection followed by spatio-temporal feature extraction. A rotation invariant spatio-temporal local binary (RIST-LBP) pattern is proposed to extract dynamic texture of the moving crowd. Further, a multi-class support vector regression is adopted for density classification. We also include a tracking step which tracks the selected interest points over the video frames for crow flow estimation. We validate our proposed approach on three different datasets such as PETS, UCF and CUHK which vary in density ranging from low to very dense. The performance of our proposed approach is compared with most commonly used pixel based statistics. Our approach has the advantage of low computational complexity with high efficiency in real world applications of video surveillance.