基于多特征和级联分类器的行人检测算法

Pedestrian detection algorithm with multiple feature and cascade classifier

针对静态图像中的人体检测问题,文章提出一种由粗到精的级联分类器人体检测算法,并改进多尺度方向(multi-scale orientation,简称MSO)特征和多尺度梯度方向直方图(Multi-scale Histograms of Oriented Gradients,简称Multi-scale HOG)特征。粗分类器采用扩展的MSO(extended multi-scale orientation,简称EMSO)特征和Adaboost级联训练得到,精分类器采用基于WTA(winner-takes-all)hash编码的Multiscale HOG(WMHOG)特征和相交核支持向量机(intersection kernel support vector machines,简称IKSVM)级联训练得到。在法国国家信息与自动化研究所(INRIA)和TUD-Brussels公共测试集上的实验结果表明,文中所提出的方法检测速度和检测率与当前代表性人体检测算法相比均有明显提高。

A coarse‐to‐fine cascade detector is proposed for the human detection problem in static images ,which uses extended multi‐scale orientation（EMSO） feature and multi‐scale Histograms of Oriented Gradients（multi‐scale HOG） feature based on winner‐takes‐all（WTA） hash .The coarse level detector employs EMSO and the Gentle Adaboost（GAB） cascade training ;the fine level detector applies multi‐scale HOG feature based on WTA hash encoding and intersection kernel support vector machines（IKSVM） cascade training .The results of the experiment on the INRIA and TUD‐Brussels public test set show that the presented method remarkably outperforms the current human detection algorithms in both detection speed and detection rate .