基于计算机视觉的行人交通信息智能检测理论和关键技术研究

发布时间：2018-09-01 10:29

【摘要】：随着视频监控系统的普及和视频图像处理技术的进步,以计算机视觉技术为基础的智能交通系统应用研究越来越受到重视,它综合利用图像处理、模式识别、人工智能等技术对监控系统采集到的视频图像序列进行处理和分析,智能化地理解视频内容并做出处理,能处理诸如事故信息判断、行人和车辆分类、交通流参数检测、运动目标跟踪等各种问题,促使智能交通系统更加智能实用,并为交通管理与控制提供全面、实时的交通状态信息。因此,基于计算机视觉的交通信息检测的方法研究具有重要的理论价值和现实意义。尽管智能视频监控技术已经研究了很多年,但基于计算机视觉的智能化交通信息检测系统仍处于发展阶段,在某些关键技术环节尚需进一步研究。目前还没有一个标准的、健壮的、精确的、高性能的目标检测和跟踪方法,还不能实时、有效的采集行人交通数据,难以智能的分析和判断行人交通的运行规律,不能对交通环境进行有效的管理和控制。在此背景下,基于计算机视觉的交通信息检测领域的研究逐步展开,并展现出良好的应用前景。本论文基于国家高技术研究发展计划(863计划)和博士科研基金项目,对基于计算机视觉技术的智能交通系统的基本理论和关键技术展开研究,结合计算机视觉的前沿研究理论,在学习使用计算机视觉开发平台Matlab的基础上,以交通视频中行人为研究对象,针对运动目标的检测、提取、跟踪、识别及交通流参数的计算与分析等问题进行了探索和研究,为ITS的智能化提供技术支持。主要研究内容包括以下几个方面：(1)首先根据计算机视觉和交通信息智能检测相关理论知识,采用图像语义层次法对行人交通语义信息智能检测过程重新进行分层,将其分成底层视觉层、中层视觉层、高层视觉层和应用层,并对每层的功能进行定义；从交通信息系统研究领域及交通信息处理流程两方面归纳和设计交通信息智能检测关键技术结构,对本文应用到的交通信息采集技术和交通数字图像处理技术进行描述；综合应用智能视频监控相关技术,构建了交通信息智能监控系统的体系结构,搭建了交通信息智能监控系统硬件和软件平台,实现理论到实践的转化,为改善和提高交通视频监控能力提供了基础。(2)根据实际交通场景中获得可靠的背景图像是比较困难的问题,提出了一种融合光流速度场的自适应背景建模方法,把光流引入背景建模中,结合背景差分结果的并运算和“死角”灰度处理来实时更新背景,实现背景建模,该模型能够精确的提取到背景图像,有效的消除噪音问题：然后,在背景拟合的基础上,提出一种基于时域和空域信息的前景目标分割方法,采用相邻多帧时域变化和Canny边缘检测法得到初始检测掩模图像,有效的解决差值局部化和噪声问题；在提取空域信息时,引入二次重构和内外标记技术对梯度图像进行修正和分水岭变换,得到空域掩模图像,有效的提高空域分割精度,消除过分割现象；最后将时空分割结果进行融合和形态学修正,精确的提取出前景目标区域。(3)在运动目标检测部分,提出一种基于形态学连通域的行人检测和底层交通语义信息提取方法,采用形态学连通域识别法,根据连通域特征进行判别,删除不相关区域,提取出交通视频图像中运动目标个数,从而精确的提取出运动行人的底层交通语义信息,为后续工作提供数据支撑；针对遮挡状态下的行人运动特性,提出一种基于人头颜色模型和轮廓信息的行人检测方法,采用RGB和YCbCr颜色空间中的聚类情况和帧差运动信息,提取候选人头区域,利用Canny边缘检测和Hough变换进行人头定位,实现目标信息统计。(4)在运动目标跟踪部分,针对Mean Shift算法存在的问题,提出基于Mean Shift目标跟踪的改进算法。构建多线索信息融合的目标表观模型,融合行人外观、空间结构和运动等多线索信息描述目标,增强特征描述能力,提高跟踪精度；从背景和目标双重角度设定目标尺度变化区域判断准则,调整算法核窗口尺寸,克服跟踪中背景干扰；采用Bhattacharyya系数判别跟踪状态,针对遮挡丢失状态,提出一种基于四部搜索策略的行人遮挡处理方法,以重新捕获丢失目标。并基于改进的目标跟踪算法,对跟踪到的行人进行中层交通语义信息提取,包括行人位移、步行速度、加速度、轨迹等；在采集信息阶段,建立ROI区域和目标链,提出基于目标跟踪的行人计数和流量统计方法,获得ROI区域的人流信息。(5)在目标底层和中层交通语义信息提取的基础上,提出一种基于递阶遗传算法改进BP神经网络的行人识别方法,采用四级递阶染色体结构描述网络结构和参数,根据构建的HGA-BP单分类器来识别交通视频图像中运动目标的种类与数量；然后在已构建的HGA-BP单分类器基础上,基于“由粗到精”的识别思想进行级联识别,构建Cascade-HGA-BP组合分类器,在底层传递高层时,采用三分检测法,实现运动行人的最终分类识别。该方法在交通场景内行人、车辆等并存的情况下,对运动行人的识别取得了很好的效果。
[Abstract]:With the popularity of video surveillance system and the progress of video image processing technology, more and more attention has been paid to the application of intelligent transportation system based on computer vision technology. It integrates image processing, pattern recognition, artificial intelligence and other technologies to process and analyze the video image sequences collected by surveillance system, intelligently. Understanding and processing video content can deal with various problems such as accident information judgment, pedestrian and vehicle classification, traffic flow parameter detection, moving target tracking and so on, which makes intelligent transportation system more intelligent and practical, and provides comprehensive and real-time traffic state information for traffic management and control. Although intelligent video surveillance technology has been studied for many years, the intelligent traffic information detection system based on computer vision is still in the development stage, and some key technologies need to be further studied. It is true that high performance target detection and tracking methods can not collect pedestrian traffic data in real time and effectively, and it is difficult to analyze and judge pedestrian traffic rules intelligently, and can not effectively manage and control the traffic environment. Based on the National High-tech Research and Development Program (863 Program) and the Ph.D. Research Fund project, this paper studies the basic theory and key technology of ITS based on computer vision technology, and combines the advanced research theory of computer vision, learning to use the computer vision development platform Matla. On the basis of B, this paper takes pedestrians in traffic video as the research object, explores and studies the detection, extraction, tracking, recognition and calculation and analysis of traffic flow parameters of moving objects, and provides technical support for ITS intellectualization. The main research contents include the following aspects: (1) First, according to computer vision and traffic information. Intelligent detection related theory knowledge, using image semantic hierarchy method to re-layered the process of pedestrian traffic semantic information intelligent detection, it is divided into the bottom visual layer, middle visual layer, high visual layer and application layer, and define the function of each layer; from the traffic information system research area and traffic information processing process two aspects This paper summarizes and designs the key technology structure of traffic information intelligent detection, describes the traffic information acquisition technology and traffic digital image processing technology applied in this paper; synthetically applies intelligent video surveillance technology, constructs the system structure of traffic information intelligent monitoring system, builds the hardware and software of traffic information intelligent monitoring system. Software platform, realizing the transformation from theory to practice, provides the foundation for improving and improving the ability of traffic video surveillance. (2) According to the fact that it is difficult to obtain reliable background images in actual traffic scenes, an adaptive background modeling method combining optical flow velocity field is proposed, which introduces optical flow into background modeling and combines background difference. The model can accurately extract the background image and effectively eliminate the noise problem. Then, on the basis of background fitting, a foreground segmentation method based on temporal and spatial information is proposed, which uses adjacent multi-frame temporal variation. Initial detection mask image is obtained by Canny edge detection method, which can effectively solve the problem of difference localization and noise. In extracting spatial information, the gradient image is corrected and watershed transformed by introducing secondary reconstruction and internal and external marking technology to obtain spatial mask image, which can effectively improve the accuracy of spatial segmentation and eliminate the phenomenon of over-segmentation. In the moving target detection part, a pedestrian detection method based on morphological connected region and a semantic information extraction method of underlying traffic are proposed. The morphological connected region recognition method is used to distinguish and delete the uncorrelated features according to the connected region features. In the region, the number of moving objects in the video image is extracted, which can accurately extract the underlying traffic semantic information of moving pedestrians and provide data support for the follow-up work. According to the pedestrian motion characteristics under occlusion, a pedestrian detection method based on head color model and contour information is proposed, which uses RGB and YCbCr color space. In the part of moving target tracking, aiming at the problems of Mean Shift algorithm, an improved algorithm based on Mean Shift algorithm is proposed. Multi-clue information fusion is constructed. Target appearance model, which combines pedestrian appearance, spatial structure and motion information to describe the target, enhances the ability of describing features and improves the tracking accuracy; sets the criteria for judging the region of target scale change from the perspective of background and target, adjusts the algorithm kernel window size to overcome the background interference in tracking; uses Bhattacharyya coefficient to discriminate. Tracking state, aiming at occlusion loss state, a pedestrian occlusion processing method based on four-part search strategy is proposed to recapture the lost target. A pedestrian counting and flow statistics method based on target tracking is proposed to obtain the pedestrian flow information in the ROI region. (5) Based on the extraction of the underlying and intermediate traffic semantic information of the target, an improved BP neural network pedestrian recognition method based on hierarchical genetic algorithm is proposed, which uses a four-level hierarchical chromosome structure. Describes the network structure and parameters, recognizes the types and numbers of moving objects in traffic video images according to the constructed HGA-BP single classifier, and then cascades recognition based on the idea of "from coarse to fine" on the basis of the constructed HGA-BP single classifier, constructs the Cascade-HGA-BP combined classifier, and transfers the high-level traffic video images to the lower level using the Cascade-HGA-BP combined classifier. Three-part detection method is used to realize the final classification and recognition of moving pedestrians. This method achieves good results in the case of coexistence of pedestrians and vehicles in traffic scenes.
【学位授予单位】：北京交通大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41;U495

【相似文献】