Robust Tracking Using Local Sparse Appearance Model and K-Selection.pdf

Robust Tracking Using Local Sparse Appearance Model and K-Selection Baiyang Liu1? Junzhou Huang1 1Rutgers, The State University of New Jersey Piscataway, NJ, 08854 baiyang, jzhuang, kulikows@ Lin Yang2 Casimir Kulikowsk1 2UMDNJ-Robert Wood Johnson Medical School Piscataway, NJ, 08854 yangli@ Abstract Online learned tracking is widely used for it’s adap- tive ability to handle appearance changes. However, it in- troduces potential drifting problems due to the accumula- tion of errors during the self-updating, especially for the occluded scenarios. The recent literature demonstrates that appropriate combinations of trackers can help bal- ance stability and flexibility requirements. We have de- veloped a robust tracking algorithm using a local sparse appearance model (SPT). A static sparse dictionary and a dynamically online updated basis distribution model the target appearance. A novel sparse representation-based voting map and sparse constraint regularized mean-shift support the robust object tracking. Besides these contri- butions, we also introduce a new dictionary learning al- gorithm with a locally constrained sparse representation, calledK-Selection. Based on a set of comprehensive exper- iments, our algorithm has demonstrated better performance than alternatives reported in the recent literature. 1. Introduction Generative and discriminative methods are two major categories used in current tracking techniques. The gen- erative models formulate the tracking problem as search- ing for the regions with the highest likelihood [5, 7, 16, 18, 19, 25, 28, 15], either using a single appearance model or multiple appearance models [27]. Discriminative methods formulate tracking as a classification problem [3, 4, 6, 11]. The trained classifier is used to discriminate the target from the background and can be updated online. Grabner et. al. [9] proposed to update the selected features incrementally using the current tracking result, which may lead to poten- tial drifting