PGT: Proposal-guided object tracking

Abstract

We propose a robust visual tracking system, which refines initial estimates of a base tracker by employing object proposal techniques. First, we decompose the base tracker into three building blocks: representation method, appearance model, and model update strategy. We then design each building block by adopting and improving ideas from recent successful trackers. Second, we propose the proposal-guided tracking (PGT) algorithm. Given proposals generated by an edge-based object proposal technique, we select only the proposals that can improve the result of the base tracker using several cues. Then, we discriminate target proposals from non-target ones, based on the nearest neighbor classification using the target and background models. Finally, we replace the result of the base tracker with the best target proposal. Experimental results demonstrate that proposed PGT algorithm provides excellent results on a visual tracking benchmark.