基于分块特征收缩的行人检测方法

Pedestrian Detection with Block Feature Shrink

针对基于梯度方向直方图(Histogram of Oriented Gradient,HOG)特征和局部二值模式(Local Binary Patterns,LBP)特征的行人检测存在特征向量维度大、检测精度有待提高的问题,提出了一种分块特征收缩的行人检测方法。首先将样本图像划分成多个大小相同的重叠分块;然后提取各分块的HOG和LBP特征,并将两种特征融合作为分块的特征,通过该特征来训练分块分类器,根据分块分类器的行人检测精度对分块进行排序,选取检测精度较高的分块进行特征收缩;最后将特征收缩后的分块特征向量连接在一起作为最终用于行人检测的特征。在INRIA公共测试集合上的实验结果表明,该方法在降低了特征向量维度的同时提高了行人检测精度。

To improve the detection rate and decrease the high dimension of histogram of oriented gradient （HOG） and local binary patterns （LBP） features in pedestrian detection,this paper proposed a pedestrian detection method based on block feature shrink.Firstly,the sample image is divided into many overlapped blocks with the same size.Then the HOG and LBP features are abstracted from these blocks,and are fused together as those blocks feature.Next,block classifiers are trained by block features.Those blocks are sorted according to the detection rate of the classifiers.We chose the blocks with higher rate to shrink their features.Finally,the block features are connected after shrinking as the last feature used to detect pedestrian.Experimental results on INRIA test set report that the proposed method has higher detection rate and lower dimension.