【作者】 李娟；

【导师】 邵春福；

【作者基本信息】 北京交通大学 ， 交通运输规划与管理， 2010， 博士

【摘要】 交通是国民经济的命脉。而当前,我国主要运输装备及核心技术水平与世界先进水平存在较大差距,运输供给能力不足,综合交通体系建设滞后,各种交通方式缺乏综合协调,交通能源消耗与环境污染问题严峻。交通问题已成为制约经济社会发展的主要瓶颈,严重影响基础设施的运转效率。《国家中长期科技发展规划纲要(2006—2020)》明确指出交通科技面临的挑战,提出优先发展智能交通系统的战略决策。智能交通系统合考虑人、车、路与环境之间的关系,对传统交通系统进行信息化、智能化、网络化和集成化改造,提高交通系统的机动性、可达性、安全性和经济性,已成为解决城市拥堵问题,提高交通运输效率的重要途径。将计算机视觉技术应用于交通数据采集领域是智能交通系统的一个重要课题。视频检测方法是交通数据采集的重要手段,相对于传统的检测方法,它不仅安装使用灵活,维护成本低,而且能够提供更大范围的监控信息和更丰富的多维交通信息。发达国家在交通视频检测方面进行了大量的研究,但大多只考虑机动车的影响,不适合我国机动车、非机动车和行人的混合交通环境。我国的城市交通具有混合、低速、高密度的运行特征,行人是交通系统的主要参与者,保障行人安全和减少其对机动车的干扰是城市交通系统安全运行的重要前提。现有的视频检测方法不能有效获得实时行人交通数据,难以分析和判断行人交通的运行规律及其对机动车的干扰,从而不能对城市道路交通进行有效地组织和控制,不利于解决我国城市特有的交通问题。在此背景下,我国在行人检测领域的研究逐步展开,并展现出良好的应用前景。本论文以行人的交通特征为理论依据,以计算机视觉和模式识别为主要技术手段,提出了若干行人交通的检测理论与方法,并用实际交通视频对这些理论和方法进行了验证,分析了系统运行结果的可靠性,为实际应用奠定基础。本论文的主要创新性成果如下：1、针对不同的摄像机架设情况,通过分析视频流特征,提出相应的行人流量检测方法。根据交通研究的需要,建立行人交通参数检测方法,可以提取行人的步行速度、启动时间和加速度等基本的行人交通特征。2、根据摄像机倾斜架设情况下的视频流特征,提出基于运动的行人流量检测方法。在运动目标检测阶段,从背景估计和前景分割两方面改进传统的混合高斯模型,使模型可以检测到由于交通冲突而静止的目标,提高了系统的鲁棒性。在阴影消除阶段,采用HSV颜色空间模型和基于形态学的目标重构方法去除运动前景中的阴影。在运动跟踪阶段,采用Kalman滤波预测行人的可能位置,缩小了目标的搜索范围,提高了系统的运行效率。在目标识别阶段,采用BP神经网络进行交通个体分类,统计行人流量。3、根据摄像机垂直架设情况下的视频流特征,提出基于人头的行人流量检测方法。在人头检测阶段,采用基于头发和肤色的混合颜色模型,分割图像,得到候选人头区域。采用Canny算法在候选人头区域提取边缘像素点,得到目标轮廓。采用Hough变换在得到的目标轮廓中进行圆环检测来定位人头。根据设定阈值进行判断,最后确定人头。在匹配计数阶段,根据人头的压线情况,进行运动和形状匹配,统计行人流量。4、根据交通研究的需要,提出行人交通参数检测方法。在标定阶段,利用“两步法”得到摄像机的内外部参数,采用二维重建算法将图像坐标转换为世界坐标,得到目标在实际交通场景中的位置。在跟踪阶段,采用光流法对目标进行运动估计。为了提高算法的鲁棒性,采用高斯金字塔对图像进行向下降采样,使光流法可以跟踪到快速运动的目标。更多还原

【Abstract】 Transportation is the lifeline of economic development. China’s transportation equipment and core technology level still has disparity from newly industrialized countries. Transportation supply ability is insufficient, different transportation modes lack coordination. The problems of energy consumption and environmental pollution seriously affect the operational efficiency of the entire society. After all, traffic problem has become the main bottleneck of economic and social development. "National Long-term Scientific and Technological Development Plan (2006-2020)" points out the challenge of transportation science and technology. It also points out the need of giving priority to the development of intelligent transportation systems (ITS).ITS is an transportation management system which integrates data communication technology, electronic sensing technology, advanced information technology, computer technology, etc. synthetically and applies them to the whole traffic management system. It has become an important way to solve traffic congestion and improve the transportation efficiency. Traffic data collection using computer vision technology is an important part of ITS. Video sensors excel many commonly used detectors due to their competitive cost, easy installation, operation and maintenance, and their ability to monitor wide areas and capture global and specific traffic data. With the development of computer vision and pattern recognition technologies, many researchers in developed countries have diving into the researches of traffic monitoring. However, most of them are concentrated on vehicles and rarely applied to non-motorized transportation modes. So they are inapplicable for mixed traffic flows which include vehicles, bicycles and pedestrians. In this condition, it is urgent to develop pedestrian detection algorithms in China, where mixed traffic flow is the major property of traffic. At present, the researches on video based traffic data collection of pedestrians and cyclists have being progressed in our country, which have demonstrated tremendous potentials of traffic application.Video based pedestrian detection methods for different traffic states are explored in this paper. It aims at extracting some important traffic parameters using traffic video sequences, which are captured by a static monocular camera in traffic scenes. Many computer vision and pattern recognition algorithms are employed. To verify the performance of the methods provided in this paper, real-world videos are tested after introduction of every pedestrian detection model respectively.What it follows contains the detailed innovations of this dissertation:1. Investigations have been carried to the video characteristics from different angles. and corresponding pedestrian flow detection methods are established respectively in this paper. To meet the need of traffic study, a traffic parameters detection method is proposed, which can capture pedestrian walking speed characteristic, pedestrian start-up time and acceleration/deceleration characteristic etc.2. A moving-based pedestrian flow detection method is provided for tilted camera configuration. This method mainly contains five modules:moving detection module, shadow remove module, feature extraction module, tracking module, and recognition module. Moving detection module aims at segmenting regions corresponding to moving objects from the rest of the image. The essential task of motion detection is to obtain an adaptive background. In order to get such background, Gaussian Mixture Model (GMM) is used for background subtraction. The main purpose of tracking module is to identify the objects correspondences between frames. To make the tracking algorithm more robust and reduce the cost of search, a prediction and matching approach is applied. In this module, the most challenging part of this system, Kalman filter (KF) is utilized to get trajectories. In classification module, in order to identify pedestrians and bicycles, Back Propagation Neural Network (BPNN) is employed. Two other simple but effective algorithms are used to alleviate the negative impacts of shadows and occlusions.3. A head-based pedestrian flow detection method is proposed for vertical camera configuration. This method mainly contains three modules:detection rectangle configuration, human head detection and matching, where human head detection is the most important part. In this module, mixed color algorithm is utilized to locate candidate human head, and then Canny algorithm is proposed to extract head contours. Finally, Hough transform is used to locate human head by head shape features.4. To capture the main traffic parameters of pedestrian, a pedestrian detection method is presented. This method mainly contains two modules:tracking and camera calibration. In camera calibration module, internal and external camera parameters are obtained by two-step approach, and then 2-D image reconstruction algorithm is utilized to transform image coordinate to world coordinate. In tracking module, optical flow algorithm is used to track the moving objects and then the data are saved automatically.更多还原