Learning Directed Intention-driven Activities using Co-Clustering

We present a novel approach for discovering directedintention-driven pedestrian activities across large urban areas.The proposed approach is based on a mutual informationco-clustering technique that simultaneously clusterstrajectory start locations in the scene which have similardistributions across stop locations and vice-versa. The clusteringassignments are obtained by minimizing the loss ofmutual information between a trajectory start-stop associationmatrix and a compressed co-clustered matrix, afterwhich the scene activities are inferred from the compressedmatrix. We demonstrate our approach using a dataset oflong duration trajectories from multiple PTZ cameras coveringa large area and show improved results over two otherpopular trajectory clustering and entry-exit learning approaches.