Improved BACF Tracker Based on Multi-Peak Detection in Visual Tracking

Correlation Filter(CF) based trackers are currently ranked top in terms of their performances on different benchmark datasets, while maintaining high frame rates. A significant amount of recent research focus on the application of strong features or discriminative model to improve comprehensive performance of the algorithm. However, lack of pertinence in some particular scenarios, especially occlusion in which most tracking algorithms can hardly well handle. The proposed tracker deals with the key causes behind the drift of tracking model under occlusion with the aim of improving performance in this scenario. By introducing Multi-Peak Detection into BACF to determine the moment of the target reappearance, the actual target can be re-detected at two positions with similar response values. In addition, this paper proposed a confidence criterion to determine the real position of the target. The proposed tracker is extensively evaluated on OTB Benchmarks: OTB-2013 and OTB-2015, and achieves an absolute improvement of 3.4% and 4.6% respectively compared to the BACF in overlap. The experimental results show the excellent performance against the state-of-the-art CF based trackers under occlusion.