Texture-based Homogeneity Analysis for Crowd Scene Modelling and Abnormality Detection

Video-based crowd behaviour analysis techniques aim at tackling challenging problems such as detecting abnormal crowd behaviours and tracking specific individuals from complex real life scenes. In this paper, an innovative spatio-temporal texture-based crowd modelling technique and its corresponding pattern analysis methods have been introduced. Through extracting and integrating those crowd textures from live or recorded videos, the so-called homogeneous random features have been deployed in the research for behavioural template matching. Experiment results have shown that the abnormality appearing in crowd scenes can be effectively and efficiently identified by using the devised methods. This new approach is envisaged to facilitate a wide spectrum of crowd analysis applications in the future through laying a solid theoretical foundation and implementation strategy for automating existing Closed-Circuit Television (CCTV)-based surveillance systems.