车辆图像稀疏特征表示及其监控视频应用  

Sparse Representation of Vehicle Image and Its' Application in Surveillance Video

作　　者：陈湘军[1,2] 阮雅端[1] 陈启美[1] 叶飞跃[2] 

机构地区：[1]南京大学电子科学与工程学院,南京210046 [2]江苏理工学院计算机工程学院,江苏常州213001

出　　处：《北京邮电大学学报》2016年第B06期81-86,共6页Journal of Beijing University of Posts and Telecommunications

基　　金：国家自然科学基金项目(61472166,61105015);江苏省科技厅项目(BE2011747);常州市应用基础研究基金项目(CJ20120021)

摘　　要：针对传统车辆图像特征在复杂场景下响鲁棒性和泛化能力低的问题,提出了车辆图像稀疏特征表示方法,并实现了基于稀疏特征的车辆图像支持向量机线性分类器,构建了基于稀疏特征和背景建模的监控车辆分类识别应用框架.与传统方法相比,该方法将车辆图像表示成字典集的低维稀疏线性组合,提高了特征表示泛化能力,能适应实时性监控视频分析的需求.实验结果表明,基于稀疏特征的车辆识别准确率比传统方法明显提升,并在低分辨率、阴影、遮挡等复杂场景下有较好的鲁棒性.Typical vehicle image feature will lost robustness and generalization ability under complex scene. To deal with this problem,sparse based vehicle images feature representation was introduced and a linear vehicles support vector machine classifier based on the sparse representation was proposed.Then,a framework of vehicle classification and recognition on surveillance video was constructed based on the background subtraction and sparse represented feature. Compared with traditional methods,vehicle images are represented as linear combination of the sparse coefficient of a learned dictionary（ atom or base） in low dimension in our method,and sparse represented feature gains higher generalization capability with less computational complexity. Experiment shows that this work exhibits better classification accuracy and robustness under complex real environment with decrease image quality of low resolution,shadow and occlusion.

关 键 词：特征表示 稀疏学习 车辆分类与识别 鲁棒性与泛化性 智能交通系统 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]