【作者】 刘光耀；

【导师】 顾伟康；

【作者基本信息】 浙江大学 ， 通信与信息系统， 2005， 博士

【摘要】 计算机视觉技术是实现图像理解的重要技术,在智能交通系统(ITS)中具有广泛的应用。基于视频图像的交通信息检测具有检测信息全面、监视范围广、成本较低、不影响交通等优点。与其他技术相比,视频技术在处理车辆跟踪、车辆分割、车辆排队等方面,具有明显的优势,是实现交通信息采集的最主要技术手段之一。在“十五”科技攻关项目“交通信息采集设备开发”课题的支持下,本文研究了固定相机条件下的交通流量参数提取、交通场景理解的技术和方法,以及基于嵌入式系统的研发工作,并成功研制一体化和分离式两种类型交通信息采集设备。 运动车辆的分割、跟踪和识别是实现交通信息检测的基础,是实现交通场景理解的关键技术。考虑到检测设备的检测功能要求和实时性要求,以及车辆形状和运动方式等有关的先验知识,本文提出了全新的基于立方体模型的车辆描述和识别系统,内容涵盖了实现一个对象识别系统所必须研究的底层处理、模型化和匹配三个方面。并对相机标定、运动信息检测和立方体模型识别等提出了行之有效的算法,最终实现了具有实用意义的检测系统。 在相机标定算法研究中,根据车辆运行于道路平面的约束知识,将相机标定分为外参数标定和内参数标定两个不同的处理步骤,实现了简化标定的目的。研究了道路平面运动约束下图像平面和道路平面之间的重投影关系,在此基础上推导出路面重投影速度与车辆运动方位之间的同方位关系,作为车辆模型定位的初始化信息。研究基于单帧图象的小波变换、层次图象分解、多尺度边缘检测、边缘描述、多尺度角点检测等技术,以及由边缘特征实现的基于平行四边形的感知聚类方法,为模型匹配提供更高层次的特征结构。 运动信息检测由变化检测和光流检测两个阶段实现。变化检测由时间差分和背景差分两种不同的差分技术实现,时间差分通过连续帧的“与”运算消除“虚影”,由背景差分实现完整的变化检测。研究了背景模型、背景更新、自动光圈抑制、形态学处理和团块凸分割等有关内容,实现稳定鲁棒的运动团块检测。光流检测由团块粗跟踪和Lucas-Kanade算法精确匹配实现,由RANSAC实现了2D仿射运动模型鲁棒估计,由Kalman滤波。此外,还阐述了基于颜色空间变换的阴影检测方法。 车辆模型由立方体结构实现模型的点、线、面之间的拓扑结构描述,以及基于视面图的2D投影描述。在识别过程中,由运动信息确定车辆模型位置和方位,对于单个车辆,实现基于轮廓的单个车辆特征检测方法,由Bayes推理网络进行识别。对于多个车辆,由边缘特征和运动光流特征,采用多种约束条件控制匹配搜索的深度和识别树的数量,完成车辆对象的分割识别。 论文最后介绍了交通流量检测设备的产品化开发,包括硬件系统和软件系统的开发。此项产品再经过一系列的改进和优化,将具有非常广阔的应用前景。更多还原

【Abstract】 Computer Vision is used to understanding the content of images, which are widely used in Intelligent Traffic System (ITS) . The video based traffic information detecting technology has detecting more information, larger surveillance areas, lower maintenance cost with respect to other sensors. The video detecting method is superiority over other technologies in vehicle tracking, vehicle detecting, and queue detecting, etc. It is the most important technology for traffic information detection. Supported with the project of Research for 10th Five-Year Plan -Development of Traffic Detection Device, this paper has researched a novel traffic flux detection and traffic scene understanding theory and technology with still cameras, with the research of embedded system, and successfully finished integrative and discrete traffic information detecting devices.Vehicles segmenting, tracking and recognizing is the basis of traffic information detection, it is the key technology for understanding traffic sense. Thinking about the detection device functional and real-time requirement, with knowledge of vehicle shapes and running style, this paper propose a new cubical model-based vehicle description and recognition system, which content includes bottom-layer features detection, object modelize, and object recognition. We develop the efficient algorithms for camera calibrating, motion detecting, model-based object recognizing, etc. We have successfully designed two kind embedded traffic information detection devices.In the research of camera calibration algorithm, with knowledge that the vehicle is running on the road-plane, the algorithm is developed into two stages - extrinsic parameter and intrinsic parameter estimating, that can reduce the .difficulty of camera calibrating. This paper researches the relation between road-plane restriction and road reprojected velocity, inferring the vehicles have the same direction to road reprojected velocity, which is used to initialize vehicle’s locations in recognize stage. This paper researches processing algorithms of single frame image, such as wavelet transform, pyramid decomposition, multi-resolution, edge description, corner detection, and parallelogram based edges perception organization.Motion information detection algorithm is separated into change detecting and optical flow detecting for two stages. Change detection is used background different and temporal different operator. The temporal difference result is used "AND" operator to remove "GHOST’, used background difference to complement the complete moving region. This paper research the background model, background updating, AGC compensation, morphologic processing and block convex segment, etc. The optical flow detection is used block matching for coarse detection and Lucas-Kanade algorithm for fine detection, used RANSAC for 2D affine robust motion model estimating. Furthermore, it describes the color-space transformed shadow detection methods.Vehicle model is described the topological relationship of cube by vertexes, lines, planes, with the aspect view of 2D reprojected image. In the recognizing stage, model is located by motion information. For a single vehicle, it is recognized by the region’s silhouette feature,inferring by Bayes networks. For mixltiple vehicles, they are segmented and recognize by edge features and optical flow information, with multiple geometry restriction to control the search depth and numbers of recognize trees.Finally, this paper describes the designing and development of traffic information detection device, with hardware system and software system.更多还原