【作者】 于海滨；

【导师】 刘济林；

【作者基本信息】 浙江大学 ， 通信与信息系统， 2007， 博士

【摘要】 人体检测与跟踪是视觉人体运动分析的重要组成部分，在视频会议、医疗诊断、高级人机交互、智能视频监控、虚拟现实以及基于内容的图像存储与检索等方面都具有广泛的应用前景和潜在的经济价值。在人体检测与跟踪技术的部分应用领域中，如智能监控、客流检测等，由于图像采集设备的安装位置受到限制或为了尽可能避免人体间的相互遮挡等原因，只能获取检测区域的俯视图像(Vertical View Image)。在这类图像中人体的各个部位当中只有头部(尤其是头顶部)体现的较为完整，头部的俯视信息(如人体头顶部轮廓，头顶部区域颜色分布等)是唯一可以用来区分多个人体目标的特征。因此，头部特征的选择与提取就成为俯视图像中人体检测与跟踪的关键环节。本文紧密围绕复杂环境中的俯视图像人体检测与跟踪问题提出了借助局部人体运动估计人体整体运动的基于头部特征提取的局部人体检测与跟踪方法。同时为了在低端嵌入式平台上实时获取人体头部特征，提出了基于改进Hough变换的头部轮廓特征提取方法和基于目标视差获取的头部深度和透视特征提取方法，并将它们用于实现具有较高准确率要求的实时人体检测与跟踪系统。为了采用复杂度较低的算法准确的提取出俯视图像中的人体头部轮廓特征，最大程度的降低误识别和漏识别的可能性，本文在对目前工程中使用最为广泛的两种改进Hough变换进行比较后选取借助边缘梯度方向对圆心轨迹进行映射的GHT方法作为提取头部类圆轮廓的主要手段。同时，为了进一步降低GHT的时间消耗并且使得GHT能够提取出与人体头部轮廓曲线拟合最好的圆形轮廓，在保留GHT参数空间映射机制的基础上对GHT的参数空间累积过程和候选圆的确认过程进行了进一步的改进，提出了针对形变较大的头部类圆轮廓检测的TGHT算法和基于视知觉分组理论的最优拟合轮廓提取方法。算法的性能测试以及实验结果均表明基于TGHT的头部轮廓特征提取方法兼顾准确性与实时性，可以在有较高准确率要求的人体检测与跟踪领域获得实时的应用。针对基于头部轮廓特征提取的单目视觉头部识别方法面临的无法从多个与头部相似的类圆区域中正确的区分出所有头部区域的困难，本文在头部轮廓特征提取的基础上采用目标视差法获得候选头部区域的视差，并将头部区域视差与深度的对应关系以及头部区域的尺度与视差的透视比例关系作为头部区域的3D特征对候选头部区域进行确认判决以进一步提升头部识别的准确率。实验结果表明，由于去除了大量与头部区域相似的类圆区域形成的虚假头部区域，因而与基于头部轮廓特征提取的单目视觉头部识别方法相比，基于头部深度和透视特征提取的立体视觉头部识别方法具有更高的头部识别准确率。为了在头部识别基础上利用运动检测和基于头部特征提取的局部人体跟踪方法完成俯视序列图像中的人体检测与跟踪，本文在基于头部特征提取的单帧图像头部识别方法基础上提出基于序列图像运动检测的边缘背景减法和基于Kalman预测以及头部区域轮廓特征帧间匹配的头部跟踪算法。本文最后依据采用的头部特征提取方法的不同提出了基于低端DSP平台的两种嵌入式视觉人体检测与跟踪系统的实现方案，并分别给出了基于头部轮廓特征提取的单目视觉人体检测与跟踪系统应用于公交客流检测现场实验的实验结果和两种基于头部特征提取的视觉人体检测与跟踪方法应用于相同的仿真图像序列的比较结果。更多还原

【Abstract】 The detection and tracking of human body is an important part of the visual analysis of human movement. It has a good prospect of application and potential economic value in video conference, medical diagnosis, interaction between man and machine, intelligent video monitoring, virtual reality, image storage and retrieval based on content and so on. In some application fields of the detection and tracking of human body, such as intelligent monitoring and passenger flow detection, only the vertical view images can be captured because of the restricted position of image capture device or to avoid occlusion among the human body as much as possible. In this kind of image, only the head, top of head in particular, will be shown completely among all of the human body parts and the information of heads in vertical view image, such as the contour of top of head and the color distribution in the head region and so on, is the only feature to distinguish the multiple human bodies from complicated circumstance. So the selection and extraction of head feature is the key of the detection and tracking of human body in vertical view image. Centering about the problem of the detection and tracking of human body in vertical view image and complicated circumstances, this dissertation presents a method to detect and track local human body based on head feature extraction and the global human body’s motion will be estimated by using the detection and tracking of local human body. To obtain the head feature in a real-time low-end embedded platform, the head’s contour feature extraction method based on modified Hough Transform and the bead’s depth and perspective feature extraction method based on target disparity acquirement are presented. And these two methods are used to realize real-time vision-based human body detection and tracking system with higher demanding of accurate rate.To extract the head’s contour feature in vertical view image exactly with lower complexity algorithm and avoid false and leak recognition as much as possible, this dissertation adopts GHT as the main technique to extract the head’s quasi-circle contour. At the same time, to further lower the time consume of GHT and to assure the head contour curve’s best fitting circle contour can be extracted, further modifications are introduced in the accumulation process in parameter space and the validation process of candidate circle on the basis of preserving the GHT’s mapping principle in parameter space, i.e. the TGHT algorithm which aims at head’s quasi-circle contour detection with larger distortion and the best fitting contour extraction method based on theory of perceptive grouping are presented. The TGHT’s performance test and the experiment results show that the head’s contour feature extraction method based on TGHT takes into account both the accuracy and the real-time performance and it can be applied in the field of real-time vision-based detection and tracking of human body with higher demanding of accurate rate.To solve the problem that head recognition method based on head’s contour feature extraction can not distinguish true head regions from many false regions similar to true head region, this dissertation obtains the stereo disparity of candidate head region by using target disparity method on the basis of bead’s contour feature extraction and the candidate head regions are validated to improve the accurate rate of head recognition by using the head’s 3D feature composed of the correspondence between head’s depth and disparity and the perspective relationship between heM’s disparity and scale. The experiment results show that the head recognition method based on bead’s depth and perspective feature extraction has higher accurate rate than the method based on head’s contour feature extraction because most false head regions have been eliminated by using the head’s 3D feature.After head recognition, to perform the detection and tracking of human body in vertical view image sequence by using the motion detection and local human body tracking based on head feature extraction, this dissertation presents two methods, i.e. the edge background subtraction method to eliminate the background edges which are out of the moving human body and the head tracking method based on Kalman prediction and head’s contour feature matching between adjacent frames.At the end of this dissertation, according to different head feature extraction methods, two kinds of embedded vision-based human body detection and tracking system based on low-end DSP are introduced. The field test results of the vision-based human body detection and tracking system based on head’s contour feature extraction applied to bus passenger flow detection and the comparison results of two kinds of vision-based human body detection and tracking method applied to the same simulation sequence image are both presented.更多还原