Efficient Online Vehicle Tracking for Real—Virtual Mapping Systems

Multi-object tracking is a vital problem as many applications require better tracking approaches. Although learning-based detectors are becoming extremely powerful, there are few tracking methods designed to work with them in real time. We explored an efficient flexible online vehicle tracking-by-detection framework suitable for real-virtual mapping systems, which combines a non-recursive temporal window search with delayed output and produces stable trajectories despite noisy detection responses. Its computation speed meets the real-time requirements, whereas its performance is comparable with that of state-of-the-art online trackers on the DETRAC dataset. The trajectories from our approach also contain the target class and color information important for virtual vehicle motion reconstruction.