【作者】 石磊；

【导师】 杨静宇；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2010， 博士

【摘要】 随着计算机电子技术与自动控制技术的发展,智能化移动平台越来越广泛地应用于军事、民用和科学研究等诸多领域。自主式车辆作为20世纪伟大的发明之一,也日益成为各国高科技战略研究的目标之一。同时,自主式车辆技术的发展促进了自动控制、模式识别、智能系统集成、传感器融合等多种学科以及信息科学技术的发展。基于计算机视觉的道路环境理解技术的研究是自主式车辆的关键技术之一,也是智能技术发展中具有挑战性意义的课题之一。尽管自主式车辆技术经过众多研究者们的深入钻研,然而许多问题仍然没有得到很好的解决,其主要原因来自于环境复杂度以及相关环境干扰因素的增加。针对这些情况,本文针对较多的道路场景进行了深入研究,主要取得的创新性研究成果包括：扩散区域Hough方法道路检测；非结构化道路环境中基于形状模型的模糊聚类分割方法(SMFCM);基于粒子群算法的道路检测新算法；以及视觉动态模型为基础的道路跟踪算法等。具体情况如下：本文针对结构化道路环境的特征,提出了一种在全局特征信息层面,基于道路边界扩展区域Hough变换的道路识别方法。对于道路环境中干扰的因素,该方法融合了道路边缘的形态特征和扩散区域在Hough空间中的全局辅助性信息。同时在实验中利用其它传感器融合模块的道路方向估计参数和部分先验的知识降低计算的复杂度,提高应用的实时性。大量的实验证明该方法在结构化和非结构化道路环境中,对道路边缘的识别有较好的作用,对于克服道路环境中干扰(阴影、不规则光照等)因素具有良好的效果和实用价值。针对非结构化道路环境的理解,本文提出了一种新的SMFCM算法。该算法利用了道路图像中道路几何的结构特征,构造了基于形状的关系隶属度矩阵,将传统的模糊聚类算法改进成了基于形状模型的模糊聚类算法。该SMFCM算法中的某些参数与道路图像环境相关。通过对较多干扰道路环境的实验,证明了该算法改进了非结构化道路环境中道路分割的效果,具有应用与研究价值。为了更精确地检测道路边缘,减少光照阴影、边缘信息模糊等造成的影响,本文提出了一种基于粒子群算法的道路检测新方法。该方法以直线变形模型为基础,在先验知识的辅助下定义了后验概率分布描述道路结构特征,利用粒子群优化算法搜索最优解。同时,针对特殊道路图像设定感兴趣友谊区域,以减低问题计算的复杂度。大量的实验证明本文算法实现了较为准确的道路边缘或者行道线的检测,能够有效地减少噪声因素的影响。在车辆运动连续性特征的基础上,结合道路环境与观测视觉的关系,本文提出了一种基于视觉动态模型的道路检测算法。首先,在道路边缘相互平行的假设条件下,仍然引用了直线变形模型来近似道路几何结构,将道路检测问题等价为一个基于先验知识的最大后验概率问题。其次,利用摄像机观测模型,车辆的状态与图像平面的模型参数之间的关系得到建立。同时在自主式车辆运动环境的基础上,建立关于运动状态的动态模型。处理过程中,由于后验概率密度函数的非凹性,使用了粒子群优化算法对采集到的第一帧图像的模型参数进行估计,将其结果作为后续动态模型初始化条件。利用非线性、非高斯等优点,使用了粒子滤波方法对连续图像序列中道路形状以及车辆的状态进行回归估计。该方法的提出能够有效地处理道路边缘特征信息不强、行道线有较多不连续性等状况,大量的实验证明该方法具有较好的应用价值。更多还原

【Abstract】 With the the development of computer electronics and autocontrol technology, intelligent moving systems are more and more widely applied to military affairs, civil utilizes, scientific fields and so on. As one of the 20th-century great inventions, autonomous vehicles are also becoming the research object for strategic high techniques all over the world. Meanwhile, they have promoted the development of many subjects such as Pattern Recognition, Intelligent Systems Integration, Sensors Fusion. As one of the key orientation, Computer Vision based road surroundings apperceiving technology is the challenging problem for intelligent technologies as well. Although many efforts have been put on Autonomous-Vehicle technology by lots of researchers, there remain some questions unsolved. The main reason is due to the increasing complexity and interfering factors for the environment. In terms of the above situation, deep research and study are carried out for many different road circumstances in this thesis, including Diffused Region of Hough method, Shape-Modeling FCM on Unstructured Roads, New Approach to Road Detection Based on Particle Swarm Optimization, and Vision Dynamic Modeling based on Road Tracking and Detection. They are introduced in details as follows:According to the characteristics of structured road environment, a Diffused Region based Hough Algorithm at global information level is proposed. As for the interfering factors of the road surroundings, the method incorporates road edge modality hint and general assistant information for the diffused region in Hough space. Meantime, Orientation estimation from other sensors and knowledge from previous experience could be utilized to decrease the time and computation complexity in experiments. Experiments specify the favorable performance on road edge identification and the effective immunity to interference with regard to structured and unstructured circumstances.Concerning the unstructured environment understanding for different roads, our thesis introduces a SMFCM (Shape-Modeling FCM) algorithm, which combines road geometry feature for the sampled images. This algorithm structures shape trait based membership matrix, which improved as shape-modeling fuzzy clustering algorithm. Besides, some parameters of the SMFCM are connected with the road scenes. Through experiments on many road images with more interference, this method is proved as preferable performance on unstructured roads and the promising value in actual applications.To get more accurate detection and understanding on road edges, and reduce the bad effect caused by shadows, blurred edge features, etc., a novel method is put forward. This algorithm is based on line deformable model, and defines a posteriori likelihood matching function with Priori information. Particle Swarm Optimization (PSO) is also applied to search the optimal solution in feature space. Furthermore, a friendly region associated with images is defined to decrease computation difficulties and other interfering factors. Experiments show the satisfactory impression for edge or lane recognition, and the ability to avoid the bad infection led by noise.Vision Dynamic Modeling based road tracking and identification mainly integrates the relation of road observation, and our thesis introduces a new algorithm with the continuity characteristic of object motion. Firstly, under the hypothesis of road edges being parallel, line deformable model is used to depict the road structure and road recognition is viewed as the problem of posterior likelihood probability all the same. Secondly, observation model for the camera is utilized to construct the relation between the vehicle state and model parameters in the image plane. Meanwhile, dynamic model for motion state is established in base of the vehicle environment. During the process of the algorithm, due to the non-concavity of the posterior likelihood function, PSO could be used to estimate the model parameters for the sampled first frame image, of which the results are employed as the initialization conditions for the motion model. With the non-linear and non-Gaussian merits, Particle Filtering (PF) is applied to reckon the road shape and vehicle state recursively. This method is proved to settle efficiently the road recognition under the condition of faint edge feature and discontinuous lanes, etc.更多还原