客运车辆危险行驶状态机器视觉辨识系统研究

发布时间：2018-06-16 18:43

本文选题：机器视觉 + 双核并行DSP　；参考：《长安大学》2013年博士论文

【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。
[Abstract]:Along with the rapid development of the highway transportation industry in our country, the problem of road traffic safety is becoming more and more prominent. The highway passenger traffic accidents are usually fatal and serious accidents, which not only cause huge economic losses to the transportation enterprises, but also make a very bad social impact on the local highway transportation management department, and even become a new society. Will the unstable factors. Therefore, the research on machine vision identification system to carry out the passenger vehicle danger, help to improve the safety of our country's highway passenger transportation and improve the ability of highway passenger traffic safety management, and after the traffic accident liability provides visual evidence, has broad application prospects and market demand.
Based on the "11th Five-Year" National Science and technology support program (2009BAG13A07) and the National Natural Science Foundation (51278062), the multi-disciplinary theory of computer graphics, information engineering, vehicle engineering, traffic engineering, and vehicle CCD visual sensing acquisition technology in machine vision technology, and embedded dual core are used in this paper. The high-speed DSP digital image processing technology, edge shape detection and analysis technology, machine learning technology and pattern recognition technology, through a large number of simulation experiments, data analysis, theoretical modeling and programming, can be used to collect real-time visual image information of passenger vehicle running state, and identify the potential of passenger vehicle in the process of running on line. A machine vision identification technology and its implementation system for dangerous driving state of passenger vehicles timely warning and recording abnormal driving behavior of drivers.
In view of the visual perception of the running state, the running track and the road environment of the passenger vehicle, the multi target feature set method is adopted to carry out the research on the identification of the road marking line and the line type and the lateral yaw warning technology of the vehicle. By strengthening the gray balance of the road image, the median filtering is quickly reorganized and the Scharr filter edge signal is filtered. Information extraction, region of interest search and constrained block scan optimal threshold segmentation processing, depth mining of road edge contour information. Based on the boundary constraints such as seed point voting area constraint, polar region constraint and chain code direction constraint, the Hough transformation is improved and the orientation detection of road identification line is realized; HSI color is fused. Spatial segmentation and dynamic window search are used to identify the line pattern of road identification line, and the region constrained particle filter tracking model is introduced to improve the detection efficiency and environmental adaptability of the road identification line. The road key information is reconstructed based on inverse perspective projection transformation, and the driving trajectory of the car in the Lane plane is predicted, and the vehicle crosswise is fully considered. The risk degree is quantified in space and time domain, and the vehicle lateral deviation warning model based on self position and time domain risk is established. The warning mechanism of the system is improved and the acceptability of the system is improved.
In view of the many interference factors in the process of image recognition in front of the vehicle, the difficulty of the complex background elimination and the limitation of the single feature representation, the method of multi-scale directional feature extraction is adopted to study the image recognition technology of the target vehicle in front of the same lane. The standard search area reduces the amount of information in the processing of the system. Through the analysis of the abrupt change characteristics of the mean value of the road surface, the existence hypothesis of the vehicle ahead is proposed. The multi scale direction feature of the vehicle gray sample is extracted with the dual channel Gabor filter, and the learning and training classification of the extracted feature samples is made by the fusion of Adaboost classifier. The location of the square vehicle in the image; based on the entropy normalization of the symmetry measure to verify the existence hypothesis of the vehicle ahead and eliminate the false targets; through the off-line training of the vehicle characteristic samples and the machine learning method combined with on-line detection, the fast, accurate recognition and positioning of the vehicle ahead are realized. The fusion improved GM (1,1) grey prediction is achieved. In this model, a small amount of historical data is used to dynamically predict the motion trajectory of the vehicle ahead, and a detection and tracking feedback mechanism is established with the interframe continuity as a clue, which alleviated the contradiction between the robustness and the real-time performance of the target vehicle detection process.
On the basis of the vehicle image recognition and location ahead, the method of human vehicle road multi source information fusion is used to study the safe distance warning technology. Through the study and analysis of the principle of monocular vision distance measurement, the single vision based on the lane plane constraint is established on the basis of the accurate calibration of the key distance parameters of the CCD vision sensor. The longitudinal vehicle distance measurement model is used to realize the accurate measurement of the longitudinal distance. Considering the driver's cognitive response characteristics, the vehicle response characteristics and the road environment, the multi-sensor information fusion technology is used to obtain the driving state information of the car and the car, and a safe distance model based on the fusion of human vehicle and road multi source information is established. The driving person emergency response probability agent, the relative driving state agent of the front car and the self vehicle and the road environment constraint agent are used as the framework, and the safe distance warning model of the group agent is established, and the early warning decision is carried out through fuzzy integral and fuzzy measure, which fully considers the influence of the external uncertainty factors and ensures the driving. It is safe to take into account the capacity of the road.
The overall design and implementation of the machine vision identification system for the dangerous driving state of passenger vehicles is discussed. The embedded dual core parallel high-speed digital image signal processing DSP and the microprocessor MCU are used as the hardware development platform, and the key components of the system are selected and the overall functional modules are designed. The storage allocation and call are optimized.
【学位授予单位】：长安大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：U492.8

【引证文献】