Real-time long-term correlation tracking by single-shot multibox detection

Abstract

Long-term robust tracking remains a challenging problem in visual object tracking. Based on the popular tracking-by-detection framework, we propose an approach named real-time long-term correlation tracking by single-shot multibox detection (RLCT-SSD), in which tracking and detection work in parallel and cooperate together. Our algorithm consists of two parts: a tracking module is expected to track in real time and accurately, whereas a detection module is responsible for verifying the tracking results at intervals, and redetecting if necessary. The detector updates the tracking module when the tracking result is unreliable. We introduce a motion model on the basis of fast discriminative scale space tracker to design our tracking module. The detection module is based on the single-shot multibox detector algorithm. To further reduce the computational cost, we use the ShuffleNet as our base network of detection. The experimental results on OTB2013 and OTB2015 demonstrate that our RLCT-SSD performs favorably against most state-of-the-art trackers and achieves long-term accurate tracking running in real time.