Robust tracking-by-detection

We extend structured tracker to be robust towards false positives and short-time occlusions. We achieve that by filtering results of the detections with Robust Kalman filter. We detect and recover from incorrect detections by comparing overlap of the bounding boxes from the filter and detector. We evaluate our method on two different tracking benchmarks. Results suggest that our tracker outperforms original Struck tracker and performs comparably with the state-of-the-art.