基于HOG-CSLBP的快速行人检测算法

摘要

In view of the problem that the feature description of existing pedestrian detection method is susceptible to light change and that the feature classification algorithm is vulnerable to uncertainty factors of sample,a variable granularity HOG-CSLBP feature was proposed.The proposed feature descriptor combined the shape information of HOG descriptor and local texture information of CSLBP descriptor to construct the new HOG-CSLBP feature.The asymmetric GentleBoost algorithm was used to train the weak detector.The pedestrian detection was realized using the detector of the nested cascade structure.Experimental results on INRIA human dataset show that the proposed method can achieve superior accuracy and work in real-time running speed.%针对现存行人检测方法中行人特征描述易受周围光照影响及特征分类算法易受样本不确定性因素干扰等问题,提出一种基于变粒度HOG-CSLBP特征的行人检测算法.对HOG特征进行变粒度编码,融合纹理特征,构造出新的HOG-CSLBP特征,利用非对称GentleBoost算法训练弱检测器,利用嵌套级联结构的检测器实现行人检测.在公开INRIA行人库上的实验结果表明,该方法与其它方法相比具有检测精度高、速度快等优点,能够满足实时检测的要求.