We have proposed advanced improvements concerning the
particle sampling and the observation model. Our advanced
particle filter was tested on image sequences taken from a moving airplane and a stationary UAV. The qualitative evaluation
showed that our efficient sampling strategy increases the robustness of the tracker in case of abrupt motion while not increasing
the risk of false positive alarms. We also demonstrated the effectiveness of our adaptive observation model, which improves
the robustness of the tracker against appearance changes of the vehicles while avoiding the template drift problem.
If the appearance or motion changes are too abrupt, the tracker may still fail. Future work could include the learning
of the semantic context within the road map to allow for better spatiotemporal guiding of the tracker.