非结构化道路路边融合算法研究

【作者】 李雪；

【导师】 唐振民；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2011， 硕士

【摘要】 非结构化道路检测技术是智能车辆自主导航的关键技术之一,是目前研究的热点与难点。单一的传感器常常无法有效的感知环境信息,陆地自主车安装了多种传感器,本文以非结构化道路路边融合为研究对象,通过融合多个传感器数据,输出可靠的道路边界位置以及道路区域内的障碍物信息,保证车辆的安全行驶。本文的融合系统使用了摄像机、红外夜视仪、激光雷达三种传感器,关于道路边界主要研究了彩色图像的道路分割、基于激光雷达的路边描述,并对激光雷达数据分析处理进行障碍检测。论文研究了一种基于彩色图像的非结构化道路分割的方法,并在此基础上提取路边特征点供给融合模块。在色度空间对H,S两个分量使用大津法选取合适的阈值进行初步分割,为了避免颜色相近的天空、植被等的影响,将视觉近区域作为检测区域。将两个分割结果合理的进行与操作,再对二值图像的噪声点使用形态学、阈值面积消去、种子填充等算法处理,将道路与非道路区域完整的区分开来。对于大面积阴影影响下的非结构化道路,进行OTSU二次分类,并根据设置的道路参考区域模型和饱和度信息归为道路与非道路类,完成道路分割。对激光雷达数据的道路边界提取设计了一种方法,先使用K均值聚类将数据划分为两类,分别代表道路左右两侧的障碍以及植被,再根据SVM算法求出使分类间隔最大的隔离带,意味着激光雷达数据描述下的道路左右边界。在障碍检测上,对激光雷达数据滤波后聚类分析,根据各类的椭圆特征信息对障碍进行识别。研究了视觉传感器与激光雷达的时空融合算法。对于多个视觉传感器提取的路边特征点,进行基于传感器模糊贴近度的加权融合,使用主元分析的直线拟合方法拟合出道路边界。激光雷达边对视觉边进行约束后,使用协方差交叉(CI)算法对二者融合得到路边位置的最佳估计值。基于D-S证据理论进行多周期障碍置信度融合,判断障碍物是否存在,提高对道路可通行区域检测的准确性和实时性。更多还原

【Abstract】 Detection of unstructured road is one of the key techniques of intellectual vehicle’s autonomous navigation. It is also a popular and difficult subject to research. Since single sensor can not understand surrounding environment well, autonomous land vehicle usually employ several sensors.This paper focus on the unmarked roads edges fusion, giving out the reliable position of the road edges and obstacles in the road through sensors fusion, so as to insure the safety of the vehicle’s driving.We use three kinds of sensors:camera, infrared sensor and laser range finder. In road boundary extraction we mainly research the segmentation of color image and the roadside description based of laser radar data. At the same time we also analysis and process the laser radar data for obstacle detection.This paper studies a method of unstructured road segmentation based on road regional characteristics. On this basis extracts the roadside points and supplies them to the fusion module.In chromaticity space use OTSU algorithm to select the right threshold for preliminary segmentation on H and S components. In order to eliminate the affect of sky and grass that is similar to road in color, we choose the visual near region for detection. After the and operation of the two results, some algorithms adopted to pick out the noise points of the binary image such as morphological, threshold value area expunction, seed filling algorithm. Then the road and non-road region are completely distinguished.With the road that is influenced by large area shadows, we undertake OTSU algorithm for the secondary classification. According to the model of road region that is primarily set by expert, generate all the kinds or region to road or non-road type. So far the road segmentation is completed.In laser radar data path boundary extraction we design a novel method.Firstly k-means clustering is used to classify the data into to two kinds, each representing the obstacles or plants next to the road.Secondly use SVM algorithm to obtain the differentiate belts that make the interval biggest. Each belt is equivalent to the road boundary under laser radar description. On obstacle detection, first step is clustering analysis and the second step is using the obstacle’s elliptic feature information to recognize the real obstacle.At last this paper research the space-and-time fusion algorithm of visuals sensors and laser radar. For the roadside points of multiple visual sensors, we propose a weighted fusion base on sensor fuzzy similarity scale. In road boundary fitting principal component analysis straight-line fitting method is adopted. After the constraint of radar laser road boundary on visual road, a fusion based on Covariance Cross Algorithm (CI) between them is used to obtain the best estimate of roadside position. We adopt the fusion of believability of multiple-cycle based on D-S evidence theory so as to judge the existence of the obstacle. Thus improve the accuracy and ability of real-time for the detection of the road region that is available.更多还原