Activity Discovery Using Compressed Suffix Trees

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