基于视觉选择性注意与IHOG-LBP特征组合的行人目标快速检测  

Fast pedestrian detection method based on combinatory features IHOG-LBP and visual selective attention computation

作　　者：刘琼[1] 陈雯柏[1] 

机构地区：[1]北京信息科技大学自动化学院,北京100192

出　　处：《计算机应用研究》2016年第1期281-285,共5页Application Research of Computers

基　　金：北京市教育委员会科技发展计划面上项目(KM201411232008);北京市属高等学校青年拔尖人才培育计划资助项目(CIT&TCD201404125)

摘　　要：常规的行人目标检测方法往往以底层特征为基础,采用密集窗口扫描的分类检测模式,其计算资源开销大而难以满足快速性要求。针对此问题,研究了一种新的行人目标快速检测方法。引入视觉选择性注意计算进行目标候选区域定位,通过提取候选区域的积分有向梯度直方图IHOG(integrated histogram of oriented gradient)特征和局部二值模式LBP(local binary pattern)特征以形成组合优势,通过级联支持向量分类方式对区域内容进行分级检测,实现了快速、可靠的行人目标检测。DET(detection error tradeoff)曲线和算法运行时间表明,相比Dalal等人的方法,本方法可在保证检测率稳定的前提下,缩短五倍的检测时间,具有更好的工程应用性。Traditional. pedestrian detection methods adopted intensive window scanning and underlying primitive features, the cost of computing resources was very large, and detection speed could not well adapt the continuously developing application re- quirements. This pape：r developed a new method to solve this problem based on visual selective attention computation. Firstly, it computed visual selective attention to position possible target regions as candidate ones. Then, it extracted IHOG （integrated histogram of oriented gradient） features and LBP（ local binary pattern） features to form combinatory features of candidate re- gions. Finally,it trained a hierarchical three level SVM classifiers with different feature dimensions and increasingly precision in a ＂cascade＂ framework to detect effectively and discard some non-pedestrian background areas rapidly. Therefore, the pro- posed method realized high speed and confidence detection of pedestrian target area. DET（ detection error tradeoff） curve and the running time of algorithms show that, compared to the Dalai et al. method,the proposed method can achieve stable true de- tection rate and shorten the detection time of 5 times, thus with better engineering applicability.

关 键 词：行人检测 视觉选择性注意 积分有向梯度直方图 局部二值模式 级联分类 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]