面向分类错误率的自适应FAST算法

Self-adaptive FAST Algorithm According to Classification Error Rate

FAST算法进行特征提取时,如果阈值和半径为非最优值,会出现特征点冗余或者丢失的现象,极大地降低了特征点的提取精度.针对上述问题,本文基于AdaBoost思想,提出了AdaBoost_FAST算法.该算法采用支持向量机作为分类器,当FAST算法中的阈值和半径非最优时,将会导致分类器错误率较高.由此根据分类器错误率计算每组阈值和半径的抽样概率,当错误率越低,其抽样概率越大,所对应的阈值和半径越接近最优值.由抽样概率构成的代价函数可知,经过多次迭代后,如果错误率较小并且无明显变化,则此时选择出的阈值和半径即为最优.实验结果表明,该算法能够有效进行阈值与半径的自适应选择,减少了特征点的冗余和丢失现象,在保证AdaBoost_FAST算法实时性的同时提高了特征点提取精度.

The feature points may be redundant or missed and the accuracy is reduced greatly when the FAST（Feature from Accelerated Segment Test）algorithm is used for feature detection with the thresholds and the radius aren′t optimal.To solve the above problem,the AdaBoost_FAST algorithm is proposed based on AdaBoost idea.The SVM is used as the classifier in the algorithm that will lead to the high error rate of classifier if the threshold and the radius aren′t optimal.A function which maps a group of threshold and radius onto a sampling probability is designed according to the error rate of the classifier.The lower the error rate,the higher sampling probability is,and the corresponding threshold and radius are closer to the optimal value.According to the cost function of sampling probability,after a number of iterations,the threshold and the radius are the optimal when the error rate doesn′t change obviously.Experiment results show that not only the self-adaptation of the threshold and radius can be realized,but also the redundancy and loss of feature points can be reduced by this algorithm.The accuracy of feature detection is improved while the real-time performance of AdaBoost_FAST algorithm is ensured.