Effective Palm Tracking with Integrated Tracker and Offline Detector

Abstract

In this paper, we propose a vision based palm tracking method with three inherently connected components: i) an offline palm detector that locates all possible palm-like objects; ii) a SURF-based tracking module that identifies the tracked palm's location using historical information; iii) an adaptive skin color model and a patch similarity calculation module. The outputs from the last component can effectively eliminate false detections and decide which palm is under tracking and also provide updated information to the first two modules. In summary, our work makes the following contributions: i) an effective offline palm detector; ii) a benchmark dataset for training and testing palm detectors; iii) an effective solution to tackling the challenges of palm tracking in adverse environments including occlusions, changing illumination and lack of context. Experiment results show that our method compare favorably with other popular tracking techniques such as Camshift and TLD in terms of precision and recall.