Stable detection of expanded target by the use of boosting random ferns

Abstract

This paper studies the problem of keypoints recognition of extended target which lacks of texture information, and introduces an approach of stable detection of these targets called boosting random ferns (BRF). As common descriptors in this circumstance do not work as well as usual cases, matching of keypoints is hence turned into classification task so as to make use of the trainable characteristic of classifier. The kernel of BRF is consisted of random ferns as the classifier and AdaBoost (Adaptive Boosting) as the frame so that accuracy of random ferns classifier can be boosted to a relatively high level. Experiments compare BRF with widely used SURF descriptor and single random ferns classifier. The result shows that BRF obtains higher recognition rate of keypoints. Besides, for image sequence, BRF provides stronger stability than SURF in target detection, which proves the efficiency of BRF aiming to extended target which lacks of texture information.