基于自适应CtF DPM特征提取的快速行人检测模型  

A Fast Pedestrian Detection Model Based on Adaptive CtF DPM Feature Extraction

作　　者：徐美华[1,2] 龚露鸣 郭爱英[1,2] 殷晓文 XU Meihua;GONG Luming;GUO Aiying;YIN Xiaowen(Microelectronic R＆D Center,Shanghai University,Shanghai 200072,China;School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200072,China)

机构地区：[1]上海大学微电子研究与开发中心,上海200072 [2]上海大学机电工程与自动化学院,上海200072

出　　处：《复旦学报（自然科学版）》2018年第4期453-461,共9页Journal of Fudan University：Natural Science

基　　金：国家自然科学基金(61376028;61674100)

摘　　要：针对行人检测系统中存在的难以同时具有较高的检测率和较快的检测速度这一问题,本文提出了一种自适应Coarse-to-Fine Deformable Part Model(CtF DPM)的行人检测模型.首先,将低分辨率根滤波器特征提取得分与阈值进行比较,以确定高分辨率部件滤波器的特征提取区域;随后,在同分辨率层中引入同级约束关系,增强同层的特征相关性;最后,将该特征提取与其他多种算法在INRIA数据库中进行检测准确性测试,并与隐式支持向量机(LSVM)结合进行实际道路环境测试.理论性能和实际测试结果表明:基于自适应CtF DPM的行人检测模型在保证检测性能的同时,特征提取时间可降至十几毫秒,显著提高了检测速度.In view of pedestrian detection system,it is difficult to balance higher detection rate and greater detection speed well at the same time.Therefore,this paper proposes an adaptive pedestrian detection model called Coarseto-Fine Deformable Part Model（CtF DPM）to improve the detection speed with strong guarantee of the detection performance.Firstly,it can identify feature extraction regions of high resolution layer part filters through comparing the feature extraction scores of the low resolution root filter with the threshold.Secondly,the same level constraint relation is introduced into the same resolution layer to enhance the characteristic correlation of this layer.Finally,the feature extraction method is compared with other typical algorithms in one Dataset under actual road environment testing.The testing results show that the pedestrian detection model based on adaptive CtF DPM can significantly reduce the feature extraction time to around ten milliseconds.

关 键 词：行人检测 特征提取 CTF DPM 同级约束 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]