Advanced Particle Filtering for Airborne Vehicle Tracking in Urban Areas

Individual vehicle trajectories from airborne image sequences provide valuable input for traffic analysis. The main characteristics of the employed camera system are given by a pixel resolution between 4.5 and 12 cm and a frame rate of 2 Hz. Three problems of particle filtering for vehicle tracking are addressed as follows. First, an adaptive motion model is presented, which controls the spreading of the particle cloud in the search space of each vehicle. Second, a spatiotemporal particle guiding approach includes the context of adjacent vehicles into the tracker to increase the stability of the tracker. Third, appearance changes of the vehicles are handled by a template update strategy. An adaptive likelihood function is introduced to balance a flexible and a strict observation model. The qualitative and quantitative evaluation on the image sequences taken from an airplane and an unmanned aerial vehicle demonstrate the improved robustness of the tracker.