基于级联特征分类器的行人检测算法

为进一步提升方向梯度直方图-局部二值模式(HOG-LBP)特征融合的行人算法在检测精度以及加快融合后的算法检测速度,提出了一种基于级联特征分类器的行人检测算法。计算样本集的方向梯度共生直方图(CoHOG)特征和鲁棒局部二值模式(RLBP)特征,使用这两种特征训练两种特征弱分类器,并将两种特征融合训练CoHOG-RLBP特征弱分类器。针对算法中存在的特征维数过高导致算法检测速度慢的问题,将各特征分类器以不同数量进行级联,构建一个6级特征弱分类器组成的级联特征分类器实现对行人目标的检测,同时使用soft-NMS算法对输出的检测窗口进行融合。在INRIA行人数据集上进行实验,实验结果表明本文算法有效提高了检测的精度与速度。

Pedestrian Detection Algorithm Based on Cascaded Feature Classifier

For improving the detection accuracy of the pedestrian algorithm based on HOG-LBP feature fusion and its detection speed,a pedestrian detection algorithm based on cascade feature classifier is proposed.Firstly,the CoHOG feature and RLBP feature of sample set are calculated,then two weak classifiers are trained by using these two features,and the two features are fused to train the weak classifier of CoHOG-RLBP feature.Aiming at the problem that the feature dimension in the algorithm is too high and the algorithm detection speed is slow,each feature classifier is cascaded in different numbers,and a cascade feature classifier composed of six-level feature weak classifier is constructed to realize the pedestrian target detection.At the same time,soft-NMS algorithm is used to detect the output detection window.Experiments on INRIA pedestrian dataset show that the proposed algorithm effectively improves the detection accuracy and speed.