【作者】 王延长；

【导师】 刘济林；

【作者基本信息】 浙江大学 ， 信息与通信工程， 2013， 博士

【摘要】 视觉导航是自主车的一种重要的导航方式。然而普通视觉传感器视场较小,无法感知道路,特别是道路交叉口的全局信息,给视觉导航造成了一定的局限性。多相机全向视觉系统(Omnidirectional Multi-camera System, OMS)具有视场大、分辨率高、畸变小等优点。本文以多相机全向视觉系统作为视觉导航传感器,以Radon变换、混合高斯模型、马尔可夫随机场和图割理论为数学工具,重点研究了自主车视觉导航中的交叉口检测、交叉口结构估计与建模、转弯参考路径的规划、转弯车速控制等问题。本文的主要内容和贡献如下：1.提出了多相机全向视觉系统外部参数的标定算法。外部参数标定是确定导航系统与环境的位姿关系的重要方法。本算法首先计算外部参数的初始值,然后采用Levenberg-Marquardt算法对其进行迭代优化而得到精确的标定结果。本文还对算法的有效性和稳定性进行了分析。2.提出一种基于混合高斯模型和马尔可夫随机场的道路检测算法。以往基于概率和机器学习的道路检测方法虽考虑了帧间的联系和影响,但忽略了像素间的空间关系,容易导致过分割。本算法将道路的分割转化为图像像素的最优二类标记问题,先利用机器学习获得混合高斯模型参数,从而得到图像属于每一类标记的概率密度场,而后由马尔可夫随机场进行空间关系建模,通过图割理论计算路与非路的最优标记。3.提出了道路交叉口的参数化模型和数据结构,并研究了该模型的回归算法。在道路交叉口的自主导航中,通常需要对交叉口的形状、结构进行描述和存储。非参数化的描述方法数据量大,不利于处理和存储。因此,本文提出了一种交叉口的参数化模型。该模型采用道路红线模型和内包络线模型对交叉口进行描述,具有良好的通用性和灵活性。其中,道路红线模型反映了交叉口的边缘信息,有利于最大化地保存交叉口的结构和形状；而内包络线模型描述了自主车安全通行的区域,主要用于自主导航和路径规划,保证了行驶的安全性。4.提出了基于Radon变换的道路交叉口检测与识别算法,并实现了基于道路交叉口参数化模型的转弯参考路径的规划和车速控制。道路交叉口的检测和识别是交叉口自主导航中的一个重要问题,它决定了导航系统对道路结构的基本判断。在已有的检测与识别算法中,有些仅检测某一特定类型的道路交叉口,有些则基于较强的前提条件或假设,这些都带来了一定的局限性。本文提出了一种基于Radon变换的道路交叉口检测和识别算法,该算法可以检测多种类型的交叉口,并识别它们的结构和类型。在对交叉口建立参数化模型后,本文还计算出自主车转弯的参考轨迹和参考半径等信息,同时由车速控制算法给出了自主车转弯过程中的限制车速。更多还原

【Abstract】 The vision-based navigation is an important way of the navigation of the Autonomous Land Vehicles (ALVs). However, due to the limited Field of View (FOV), the ordinary vision sensors are unable to perceive the global information of the roads, especially the road intersections. This causes the limitations to the vision-based navigation.The Omnidirectional Multi-camera System (OMS) has advantages of large FOV, high resolution and slight distortion. Taking the OMS as the vision sensor for the navigation, this dissertation places emphasis on the study of the detection, the structural estimation and the modeling of an intersection and the planning of the referential path and the limited speed in the ALV’s navigation. The study is based on the mathematical tools of the Radon Transform, the Gaussian Mixture Model (GMM), the Markov Random Field (MRF) and the Graph Cuts.The main contents and contributions of this dissertation are as follows:1. We propose the algorithms for solving the calibration and the rectification for an OMS. The extrinsic parameters are important to determine the pose of the navigation system. Firstly, the initial guesses of the extrinsic parameters are calculated. Then the final results are refined with the Levenberg-Marquardt algorithm. We have analysed the validity and reliability of the algorithms.2. For detecting the roads, we introduce a kind of road detection algorithm based on the GMM and the MRF. Although the ordinary road detection algorithms based on the probability and the machine learning take the relation and the influence between the frames into account, they ignore the spatial relationship between the pixels, thus resulting in the over-segmentation. This algorithm treats the road segmentation as a binary classification problem. Through the GMM and the machine learning we can obtain the probability density field of the road and the non-road area of the image. By the MRF and the Graph Cuts, the optimized labels of the road and the non-road can be calculated. 3. This dissertation also puts forward the parametric model and the data structure of the road intersection and studies the regression algorithm of the model at the same time. In the ALV’s navigation in the intersection region, people often need to descript and store the shape and the structure information of the intersection. Suffering from the big data, the non-parametric methods are not suitable for the processing and the storage. To solve these problems, we propose the parametric model of the intersection. The model has a strong adaptability and flexibility, using the inner envelope and the boundary lines of the roads to describe the intersection. On the one hand, the boundary line model records the actual information of the road boundary itself. On the other hand, the inner envelope model guarantees the safety of the vehicle, mainly used for the navigation and the path planning.4. This study also presents a road intersection detection and recognition algorithm based on the Radon Transform, and implements the planning of the referential path and the limited speed based on the parametric model of the road intersection. The intersection detection is an important problem of the ALV’s navigation in the intersection region. It decides the basic judgement to the road structure. Among the existing methods of the intersection detection, some of them only detect a certain type of the intersections and some are based on a strict condition or hypothesis, thus causing some limitations. The detection and recognition algorithm in this study is able to detect multiple types of the intersection, and recognize the shape and the structure. After modeling the intersection, the algorithm also plans the referential path, the radius and the limited speed of the ALV’s turning.更多还原