Abstract
Most modern trackers typically employ a bounding boxgiven in the first frame to track visual objects, where theirtracking results are often sensitive to the initialization. Inthis paper, we propose a new tracking method, ReliablePatch Trackers (RPT), which attempts to identify and exploitthe reliable patches that can be tracked effectively throughthe whole tracking process. Specifically, we present a track-ing reliability metric to measure how reliably a patch canbe tracked, where a probability model is proposed to esti-mate the distribution of reliable patches under a sequentialMonte Carlo framework. As the reliable patches distributedover the image, we exploit the motion trajectories to dis-tinguish them from the background. Therefore, the visualobject can be defined as the clustering of homo-trajectorypatches, where a Hough voting-like scheme is employed toestimate the target state. Encouraging experimental resultson a large set of sequences showed that the proposed ap-proach is very effective and in comparison to the state-of-the-art trackers. The full source code of our implementationwill be publicly available.