Rapid target recognition and tracking under large scale variation using semi-naive Bayesian

In this paper, we present a robust feature matching-based solution to real-time target recognition and tracking under large scale variation using affordable memory consumption. In order to extract keypoints robust to scale, viewpoint changes and partial occlusions, we propose a training scheme based on FAST to detect the most repeatable features in target region. As for feature matching, Ferns suffers from unaffordable memory consumption for lower-power hardware platform, by modifying the original Ferns, we achieve comparable results with only a tiny fraction of runtime memory, which is one aspect of our contribution. To handle with long distance, large scale variation target tracking, we take advantage of multi-model tactics, which is another contribution of us. At last, a typical tracking experiment with speed over 40 fps on a 2.0 GHz PC confirms the efficiency of our approach.