基于非监督特征学习的分叉道路检测算法  

Road detection algorithm for crossroad based on unsupervised feature learning

作　　者：杨力[1] 刘济林[2] 

机构地区：[1]中国计量学院信息工程学院,浙江杭州310018 [2]浙江大学信息与电子工程学系,浙江杭州310027

出　　处：《浙江大学学报（工学版）》2014年第9期1558-1563,共6页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(60534070;90820306)

摘　　要：为了解决道路分叉环境中的智能车辆导航问题,提出一种大视场、近距离的道路检测方法.采用安装于车头的鱼眼摄像机,克服了普通相机视野窄、近处存在盲区的问题;通过鱼眼图像重投影,去除鱼眼畸变和透视失真,获得尺度一致的数据块;应用非监督特征学习和逻辑回归分类器,从海量未标记数据中得到原始数据块的稀疏表达,免除了人工标记数据,最后得到路面可通行概率.实验结果表明:此算法在缺乏先验道路几何信息、无手工标记数据的情况下,可以正确地识别分叉道路可通行区域,无视野盲区.A road detection method with large field-of-view and near range was proposed to detect the road area at crossroads for intelligent vehicle navigation. A fisheye camera mounted in the front of the vehicle was used to cover the blind spots in traditional methods. An image reprojection was applied to eliminate ra- dial distortion and perspective projection distortion, thus to get the image blocks of the same scale. Unsu- pervised features learning was adopted to get the sparse representation of original image blocks, and then road traversable probability was calculated by logistic regression classifier, with no need to label the mass data manually. Experimental results show that this algorithm can recognize the road area at crossroads without blind spots in the absence of prior geometrical information and manual labeling.

关 键 词：道路检测 特征学习 逻辑回归 视觉导航 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]