Trajectory Clustering by Sampling and Density

AbstractThe trajectory data of moving objects contain huge amounts of information pertaining to traffic flow. It is incredibly important to extract valuable knowledge from this particular kind of data. Trajectory clustering is one of the most widely used approaches to complete this extraction. However, the current practice of trajectory clustering always groups similar subtrajectories that are partitioned from the trajectories; these methods would thus lose important information of the trajectory as a whole. To deal with this problem, this paper introduces a new trajectory-clustering algorithm based on sampling and density, which groups similar traffic movement tracks (car, ship, airplane, etc.) for further analysis of the characteristics of traffic flow. In particular, this paper proposes a novel technique of measuring distances between trajectories using point sampling. This distance measure does not divide the trajectory and thus conserves the integrated knowledge of these trajectories. This trajectory clustering approach is a new adaptation of a density-based clustering algorithm to the trajectories of moving objects. This paper then adopts the entropy theory as the heuristic for selecting the parameter values of this algorithm and the sum of the squared error method for measuring the clustering quality. Experiments on real ship trajectory data have shown that this algorithm is superior to the classical method TRACLUSS in the run time and that this method works well in discovering traffic flow patterns.