Tracking multiple vehicles in airborne image sequences of complex urban environments

Airborne image sequences provide useful data for monitoring traffic scenes in large complex urban environments. Tracking vehicles in these low frame rate images is challenging, because of abrupt motion and low resolution description of the targets. In this work, we propose a new particle based method for long-term tracking of multiple vehicles in airborne images. Our tracking system uses a vehicle-specific motion model and integrates shape and color information in the observation model. The state of each vehicle is robustly estimated by a mean shift clustering method. Multiple vehicles are tracked by individual particle filters while their interaction is incorporated into the motion model. The data association problem is resolved by finding the maximum weighted matching in a bipartite graph. Additionally, the tracking performance is improved by biasing the particle positions toward the trajectory of preceding vehicles. The experimental results show that our tracker is robust against illumination changes and can handle total occlusions over several image frames. Thus, our tracker produces reliable trajectories that can be used to determine the traffic dynamics.