Dynamics Based Trajectory Segmentation for UAV videos

A novel representation of vehicle trajectories is proposedfor applications in trajectory analysis and activity detection.Specifically, a piecewise arc fitting based smoothingalgorithm is proposed for denoising the trajectories. A dynamicprogram is used to find the optimal arc fit to a giventrajectory. We motivate the usage of dynamic primitivesto parametrize common vehicular activities, and proposea dynamics based trajectory segmentation algorithm. Eachprimitive is modeled using a second order Auto-Regressivemodel, and form useful descriptors for a given vehiculartrajectory. We evaluate both our trajectory smoothing anddynamic trajectory segmentation algorithm on a real UAVvideo dataset, and show performance improvements whichclearly motivate its wide applicability in a general trajectoryanalysis system.