Extracting Pathlets FromWeak Tracking Data

Abstract

We present a novel framework for extracting “pathlets”from tracking data. A pathlet is defined as a motion regionthat contains tracks having the same origin and destinationin the scene and that are temporally correlated. The proposedmethod requires only weak tracking data (multiplefragmented tracks per target). We employ a probabilisticstate space representation to construct a Markovian transitionmodel and estimate the scene entry/exit locations. Theresulting model is treated as a set of vertices in a graph anda similarity matrix is built which describes broader nonlocalrelationships between states. A Spectral Clusteringapproach is then used to automatically extract the pathletsof the scene. We present experimental results from scenes ofvarying difficulty and compare against other approaches.