FEATURE SIMILARITY EVALUATION AND EVENT CLUSTERING FOR VISION SYSTEMS
W. S. K. Fernando, Pramuditha Perera, M. P. B. Ekanayake, Roshan Indika Godaliyadda, Janaka Wijayakulasooriya · 2015
An approach to perform video event classification by taking into account feature similarity through spectral clustering is proposed in the paper. Spectral clustering is used to identify different types of events and to define respective event signatures. Number of modifications to standard spectral clustering algorithm is proposed to make it suited for an event classification application. Dynamic Time Warping (DTW) is used to overcome the challenge of evaluating pairwise inter-event disparity values which is required for spectral clustering. Methods to determine initial parameters for spectral clustering are also proposed. Observed events are compared against event signatures to determine their event identity and to detect anomaly events. Event classification results obtained for a series of videos that involve human motion are discussed in the paper.