基于先验Haar-Like特征的多通道人体检测方法

Multi-channel human detection method based on prior Haar-Like features

针对聚合通道特征(ACF)在人体检测中对人体轮廓特征描述不够充分的问题,提出基于先验知识的Haar-Like特征来增强检测器对人体轮廓特征的描述能力。设计了一组Haar-Like特征,利用了人体上半身轮廓特点的先验知识,将头部、上半身以及背景视为3个不同的部分。实验结果表明:相比于ACF等算法,所提方法能够提高检测器检测精度,在INRIA数据集上召回率为94. 57%。

Aiming at the problem that the aggregate channel feature(ACF)does not adequately describe human contour features in human detection,a Haar-Like feature based on prior knowledge is used to enhance the ability of the detector to describe human contour features.A set of Haar-Like features is designed.This feature takes advantage of prior knowledge of the human body’s upper body contour characteristics and treats the head,upper body,and background as three different parts.The experimental results show that compared with the ACF and other algorithms,the detection method combined with the Haar-Like feature can improve the detection precision of the detector,and the recall rate is increased to 94.57%on the INRIA data set.