Classification of Object Trajectories Represented by High-Level Features Using Unsupervised Learning

Abstract

Object motion trajectory classification is an important task, often used to detect abnormal movement patterns for taking appropriate actions to prohibit occurrences of unwanted events. Given a set of trajectories recorded over a period of time, they can be clustered to understand usual flow of movement or detection of unusual flow. Automatic traffic management, visual surveillance, behavioral understanding, and sports or scientific video analysis are some of the typical applications that benefit from clustering object trajectories. In this paper, we have proposed an unsupervised way of clustering object trajectories to filter out movements that deviate large from the usual patterns. A scene is divided into nonoverlapping rectangular blocks and importance of each block is estimated. Two statistical parameters that closely describe the dynamic of the block are estimated. Next, these high-level features are used to cluster the set of trajectories using k-means clustering technique. Experimental results using public datasets reveal that, our proposed method can categorize object trajectories with higher accuracy when compared to clustering obtained using raw trajectory data or grouped using complex method such as spectral clustering.