Modes of clustering for motion pattern analysis in video surveillance

Abstract

This work introduces a set of tools for motion pattern analysis in video surveillance. For a given video stream, first the motion trajectories are extracted and an affinity matrix is constructed. Then, motion pattern analysis is conducted based on Normalized Spectral Clustering. An Eigengap based methodology is proposed for determining the number of clusters. It was observed that in real life scenarios, according to human perception, the number of clusters observed is not a global constant, that it actually can take multiple values based on the level of zooming. Thus, a novel concept called ‘Modes of Clustering’ is introduced, where ‘Modes’ correspond to the multiple clustering arrangements that exist for a given scenario. The number of clusters and the arrangement of trajectories within those clusters serve as a descriptor for each such ‘Mode’. The free parameter Sigma in the standard Spectral Clustering algorithm, can be used as a tool for zooming. Accordingly, a ‘Sigma Sweep’ is introduced as a methodology for detecting the significant modes. Hence, a more detailed and accurate representation closely reflecting human perception is proposed and its applicability for video surveillance is elaborated through a case study.