Feature Tracking in Long Video Sequences Using Point Trajectories

This article proposes a robust method for tracking sparse point features that is stable to long video sequences. To improve the stability of feature tracking in a long sequence, we use feature trajectories of tracked points. We predict feature states in the current image frame and compute rough estimates of feature locations. A search window is positioned at each estimated location and a similarity measure is computed within the window. Through the refinement using the similarity measure, an accurate feature position is determined. To reduce false matches, an outlier rejection stage is also introduced. Experimental results from several real video sequences showed that the proposed method stably tracks point features for long frame sequences.