A multiple-candidate-regeneration-based object tracking system with enhanced learning capability by nearest neighbor classifier

We present a real-time object tracking system that employs the nearest neighbor classifier and the multiple candidate regeneration as the appearance model and the searching strategy. Based on the analysis of the likelihood measurement, a novel appearance model has been proposed, with an online learning strategy specifically designed for tracking task. The number of templates is reduced to a small number and reliable template selection is realized. Because of its simplicity, the system can be efficiently built on hardware. The evaluation of the proposed system on challenging video sequences shows robust tracking capability with accurate tracking results. Hardware implementation of this system is also discussed and a processing speed much faster than the frame rate can be expected.