Online tracking and event clustering for vision systems

This paper proposes a comprehensive method for online-event clustering in videos. Adaptive Gaussian mixture model was modified to obtain consistent foreground estimates for object tracking by introducing shadow filtering, stillness handling, visual impulse removal and visual distortion filtering. Object-events were defined in terms of feature trajectories of foreground and they were modelled using the time series modelling technique. A cross-substitution based model comparison method was employed to compare the disparity between events. Spectral clustering (SC) was utilised to cluster events, and methods for SC initial parameter selection have been proposed. A method for cluster identity assignment in consecutive clustering iterations is also utilised to handle the evolving nature of the unsupervised learning methodology adopted. The proposed method is capable of producing reliable clustering results online, amidst a number of complications including dynamic backgrounds, object shadows, camera distortions, sudden foreground bursts and inter-object interactions.