Activity discovery using compressed suffix trees

摘要: The area of unsupervised activity categorization in computer vision is much less explored compared to the general practice supervised learning patterns. Recent works lines "discovery" have proposed use probabilistic suffix trees (PST) and its variants which learn models from temporally ordered sequences object states. Such often contain lots objectstate self-transitions resulting a large number PST nodes learned models. We propose an alternative method mining these by avoiding while maintaining useful statistical properties thereby forming "compressed tree" (CST). show that, on arbitrary with significant self-transitions, CST achieves lesser size as polynomial growth PST. further distance metric between CSTs using which, are categorized hierarchical agglomerative clustering. trajectories extracted two data sets clustered for experimental verification discovery.