摘要：

针对聚合通道特征(ACF)算法误检窗口多的问题,提出一种由粗到精的级联行人检测算法。采用ACF算法快速粗检,改进通道特征来滤除误检窗口,以每个图像通道学习主成分分析(PCA)滤波器组,代替PCANet从训练图像和卷积图中学习滤波器组,用图像通道进行单层卷积,代替PCANet的双层卷积以降低特征维数,提升对行人的表达能力,并对卷积图池化降维,得到改进的通道特征。仿真结果表明,该算法相对于原ACF算法误检窗口减少,检测率在INRIA、Caltech数据库上分别提高3.8%和17.5%。

关键词: 行人检测, 聚合通道特征算法, 积分通道特征, 卷积网络, 主成分分析, 自动学习

Abstract:

As the Aggregate Channel Features(ACF) algorithm has many false detection windows,a coarse-to-fine cascaded pedestrian detection algorithm is proposed.ACF is employed as the coarse detector,and then the channel features are improved to filter out false detection windows.The Principal Component Analysis(PCA) filter bank are learned from each channel,instead of learning filter bank from the training image and convolution maps.The single-layer convolution is executed on the filter bank and channels,to reduce feature dimension and improve feature discrimination capability,Finally,a pooling operation is applied on convolution maps to reduce the feature dimensionobtaining improving channel features.Simulation results show that compared with the original ACF algorithm,the proposed method has less false detection windows and the detection rate on INRIA and Caltech databases increases by 3.8% and 17.5% respectively.

Key words: pedestrian detection, Aggregate Channel Features(ACF) algorithm, Integral Channel Features(ICF), convolutional network, Principal Component Analysis(PCA), automatic learning

中图分类号: 

• TP391.41