Occlusion-Robust Pedestrian Tracking in Crowded Scenes

This paper focuses on tracking in typical traffic monitoring scenarios with emphasis on handling occlusions caused by trees, lampposts and cables. We extend the existing TRracking with Occlusion handling and Drift correction (TROD) algorithm with a novel occlusion detection algorithm, based on measuring the changes in the object motion pattern. The motion information is extracted via frame differencing and described a the HOG descriptor. Occlusions are handled by preventing the model update and predicting the object location based on prior observations. Our proposed system clearly outperforms state-of-the-art tracking algorithms for larger occlusions in the specific pedestrian surveillance scenario, that is, the percentage of successfully tracked objects grows with 10-15%. At the same time, for non-specific public datasets, the performance is similar to existing state-of-the-art tracking algorithms.