Multi-Bernoulli filtering for keypoint-based visual tracking

Abstract

In this paper, we consider a single object visual tracking problem using multi-object filtering technique. We represent object appearance as a multi-object distribution of keypoints. Hidden positions of keypoints are observed by using SURF feature detectors and multi-Bernoulli filtering is used for tracking of keypoints. Unlike other feature matching based object trackers, multi-Bernoulli filtering based tracker is free from combinatorial matching problem. The estimated number of keypoints can be used as a quality measure to determine track re-initialization when it is necessary. Experimental results show that multi-object filtering can be one of effective solutions for single object visual tracking.