【作者】 杨炜；

【导师】 魏朗；

【作者基本信息】 长安大学 ， 车辆工程， 2013， 博士

【摘要】 随着我国公路交通运输业快速发展的同时，道路交通安全问题日益突出，公路客运事故一般都是人员死伤惨重的恶性事故，不仅给运输企业造成巨大的经济损失，而且给当地公路运输管理部门造成了极坏的社会影响，甚至成为了新的社会不稳定因素。因此，开展客运车辆危险行驶状态机器视觉辨识系统的研究，有助于改善我国公路客运安全性和提高公路客运安全管理能力，并能够对发生交通事故之后的责任认定提供部分可视化证据，具有广阔的应用前景和市场需求。本文依托“十一五”国家科技支撑计划重大项目(2009BAG13A07)和国家自然科学基金项目(51278062)，综合运用计算机图形学、信息工程学、车辆工程学、交通工程学等多学科理论以及机器视觉技术中的车载CCD视觉传感采集技术、嵌入式双核并行高速DSP数字图像处理技术、边缘形状检测与分析技术、机器学习技术与模式识别技术，通过大量模拟试验、数据分析、理论建模和程序设计，研究能够实时采集客运车辆行驶状态视觉图像信息，在线辨识客运车辆行驶过程中存在的潜在危险，适时警示和记录驾驶人非正常驾驶行为的客运车辆危险行驶状态机器视觉辨识技术及其实现系统。针对客运车辆行驶状态、运行轨迹和道路环境的视觉感知问题，采用多目标特征集合的方法，进行了道路标识线方位与线型识别以及车辆横向偏航警告技术的研究。通过对道路图像灰度均衡化增强、快速重组中值滤波、Scharr滤波边缘信息提取、感兴趣区域搜索和约束块扫描式最优阈值分割处理，深度挖掘道路边缘轮廓信息。基于种子点投票区域约束、极角区域约束以及链码方向约束等边界约束条件，对Hough变换进行改进并实现了道路标识线的方位检测；融合HSI色彩空间分割与动态窗口搜索实现了道路标识线线型的辨识；引入区域约束粒子滤波跟踪模型，提高了道路标识线的检测效率和环境适应能力。依据逆透视投影变换重建道路关键信息，预测车道平面内自车的行驶轨迹，充分考虑自车横向分速率和横向偏航角的影响，在空间域和时间域内量化危险度，建立了基于自车位姿与时域危险度的车辆横向偏航警告模型，改善了系统的警告机制，提高了系统的可接受度。针对前方车辆图像识别过程中存在的干扰因素较多、复杂背景排除困难和单一特征表示的局限性等问题，采用多尺度方向特征提取的方法，进行了同车道内自车前方的目标车辆图像识别技术的研究。充分挖掘前方车辆图像信息设置目标搜索区域，减小了系统运算处理信息量。通过对路面灰度均值突变特征的分析，提出前方车辆存在性假设；利用双通道Gabor滤波器提取车辆灰度样本的多尺度方向特征，融合Adaboost分类器对提取的特征样本进行学习训练分类，确定前方车辆在图像中的位置；依据信息熵归一化对称性测度，验证前方车辆存在性假设，排除虚假目标；通过车辆特征样本的离线训练与在线检测相结合的机器学习方式，实现了前方车辆快速、准确的识别和定位。融合改进GM(1,1)灰色预测模型，利用少量历史数据信息动态预测前方车辆的运动轨迹，并以帧间连续性为线索，建立了一种检测与跟踪反馈工作机制，缓和了目标车辆检测过程中鲁棒性与实时性之间的矛盾。在前方车辆图像识别定位的基础上，采用人-车-路多源信息融合的方法，对安全车距预警技术进行了深入研究。通过对单目视觉测距原理的研究分析，在CCD视觉传感器关键测距参数精确标定的基础上，建立了基于车道平面约束的单目视觉纵向车距测量模型，实现了纵向车距的精确测量。充分考虑驾驶人认知响应特征、车辆响应特性和道路环境等因素，运用多传感器信息融合技术获取前车及自车的行驶状态信息，建立了基于人-车-路多源信息融合的安全车距模型。以驾驶人应急响应概率智能体、前车与自车相对行驶状态智能体和道路环境约束智能体互相协作为架构，建立了群智能体协作的安全车距预警模型，通过模糊积分与模糊测度进行预警决策，充分考虑了外界不确定性因素的影响，在保证行车安全的同时兼顾了道路的通行能力。探讨了客运车辆危险行驶状态机器视觉辨识系统的总体设计与实现，以嵌入式双核并行高速数字图像信号处理DSP和微处理器MCU作为硬件开发平台，完成了系统关键部件的选型以及总体功能模块的设计，并对系统图像处理过程中的内存分配和调用进行了优化设计。更多还原

【Abstract】 With the rapid development of highway transportation industry, road traffic safetyproblems have been increasingly prominent. Generally highway passenger transportationaccidents are personnel and malignant, they not only caused huge economic losses totransport enterprise, but also had a bad social influence on local road transport administration.What’s more, they have become a new social unstable factor to some extent. So carrying outhighway passenger vehicle dangerous driving status machine vision recongition systemresearch, improving highway passenger safety and road passenger transport safetymanagement ability, providing some visual evidence after the traffic accident, have wideapplication future and market demand.Relying on the "11th five-year plan" national science and technology support plan keyprojects (2009BAG13A07) and the national natural science fund project (51278062), thispaper applies a combination of computer graphics, information engineering, vehicleengineering, traffic engineering and other multi-disciplinary theories, and computer visiontechnology for on-board CCD image sensor technology, embedded dure-core parallelhigh-speed DSP digital image processing technology, features of shape edge detection andanalysize technology, machine vision and pattern recognition technology. Through a largenumber of simulation tests, data analysis, theoretical modeling and programming, study thevisual recognition system which can collect real-time passenger vehicle driving state visualimage information, and online identify vehicle driving potential dangers that exist in theprocess of driving, warning and recording passenger vehicle drivers’ improper drivingbehavior.Aiming at visual perception problems of passenger vehicle driving status, movingtrajectory and road environment, applying the method of multi-objective characteristiccollection, study the recognition of road marking lines’ position and linear and vehicle lateralyaw warning technology. Through the road image gray equalization enhancement, rapidrestructuring of median filtering, Scharr filtering extracting edge information, searching ROIarea and bound quick scanning optimal threshold segmentation, excavate road edge’s profileinformation deeply. Combining constraint of joint seed point voting area with constraint of polar angle and constraint of boundary in chain code direction, improving the Houghtransform to achieve the goal of azimuth detection of road marking line; realize linearidentification by HSI color space segmentation and dynamic window search; introduce areaconstraint particle filter for dynamic tracking and improve the detection efficiency of roadmarking line and environment adaptability. According to inverse perspective projectiontransformation, rebuild the key information of road, estimate the driving track of vehiclewhich is in the lane plane, comprehensive consideration of the influence of yaw rate andlateral angle, quantitative risk within the space domain and time domainon, establish thelateral yawing warning model which is based on the parking posture and risk of time domain,improve the warning mechanism and the acceptable degree of the system.Aiming at the problems that too many interference factors difficulties lie in ruling out thecomplex background, limitations result from single feature representation that exist in theprocess of the front vehicle recognition, carries on the research on recognition technology ofthe front car in the same lane plane by the feature extraction method of multi-scale direction.Fully excavcating the image information of front vehicle, set target search area, reducing theprocessing amount of system calculation. Through the analysis of the pavement grayscaleaverage mutation characteristics, put forward the front vehicle existence assumption; extractsthe multi-scale direction characteristics of vehicle gray sample by using dual channel Gaborfilter, fuse the extracted features which is extracted by Adaboost classifier to learn, train andclassify, detect the position in the image of the front vehicle; verifies the existence assumptionof front vehicle on the ground of information entropy normalized symmetry measure, theneliminates the false targets; realizes the detection and location of front vehicle through themachine learning method of a combination of off-line training and on-line detection ofvehicles’ characteristic sample. Fusing the improved GM(1,1) gray forecasting model,dynamically predicts the moving track of front vehicle through only a small amount datainformation. Using the continuity of frame interval as clues, establishes a detection andtracking feedback working mechanism to defuse the target vehicle contradiction betweenreal-time and robustness.On the basis of image recognition and location of front vehicle, applyingdriver-vehicle-road multi-source information fusion method to further study the safety vehicle distance recognition and warning technology. Through the theoretical analysis of monocularvision range finding principle, establishes the monocular vision vertical distance measurementmodel which is base on the lane plane constraint on the basis of accurate calibration of CCDimage sensor key measure parameters, realized the precise measurement of the verticaldistances. Given full consideration to the driver’s cognitive response characteristics, vehicleresponse characteristics and road environment factors, using multi-sensor information fusiontechnology to get vehicle running status information of front vehicle and host vehicle,establish the safety distance model which is based on multi-source, such as driver, vehicle,road information fusion. Collaborating driver emergency response probability agent withrelative state agent of front vehicle and host vehicle and road environment constraints agent asarchitecture, establishes Multi-Agent system for safety distance warning model. Given fullconsideration to the impact of outside uncertainty factors, through fuzzy integral and fuzzymeasure as the warning decision, guarantees the driving safety as well as the capacity ofhighway traffic.Discussed the overall design and implementation of machine vision recognition systemfor passenger vehicles’ dangerous driving state, with embedded dual-core parallel high-speeddigital image signal processing of DSP and microprocessor MCU as hardware developmentplatform, completed the selection of key system components and design of overall functionmodule, optimized the memory allocation and transfer for the system image processing.更多还原