【作者】 赵敏；

【导师】 孙棣华；

【作者基本信息】 重庆大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 行人检测技术在智能监控系统、智能车辆/辅助驾驶、运动分析、高级人机接口等众多方面具有广泛的应用前景和研究价值,是近年来计算机视觉领域备受关注的前沿方向和研究热点。行人检测包括了行人目标的检测、提取、识别和跟踪等方面的内容,由于应用场所各异、环境复杂,加之具有的服饰变化、姿态变化、人体运动的随意性和随机性、遮挡等方面特点,使其成为一个具有挑战性的困难问题,受到学术界广泛关注。为了克服多行人目标相互遮挡带来的影响,本文采用单目垂直摄像方式获取检测区域的行人目标图像,建立借助人体特定部位-头部的特征实现多行人目标的检测与跟踪方案。围绕形变与尺度变化下的多头部目标的检测与跟踪,重点研究了静态图像中头部目标的分割、基于多特征的头部目标识别,彩色图像序列中运动行人目标的快速检测,以及复杂场景中尺度变化的行人目标的跟踪问题,形成了一套序列图像中的多行人目标自动检测与跟踪方法。论文主要内容如下:①讨论了行人检测技术的研究背景、发展现状及研究意义,在对各种信息感知手段进行综合比较的基础上,重点阐述了基于计算机视觉的行人检测技术研究现状及研究进展,分析了当前行人检测技术在目标识别与跟踪方面的不足。②详细分析了单目视觉下行人信息的采集方式以及在不同采集方式下行人图像的特点,针对基于整体与基于部位组合的行人检测方法难以实现遮挡情况下的多行人目标的正确检测,本文将人体的特定部位-头部特征作为区分多个人体目标的特征。在此基础上,通过对垂直单目视觉下现有头部检测方法以及行人头部特征的分析,提出结合颜色特征、形状特征、运动特征等多种特征实现头部目标的检测方案。③研究了静态彩色图像中的多头部目标分割与识别技术。针对基于发色或灰度特征的头部目标分割方法易受光照的影响,难以完整地分割出头部目标区域的问题,提出一种面向头部区域候选目标提取的改进mean shift彩色图像分割算法。该算法对经典mean shift分割算法进行了两个方面的改进:一是基于不同带宽下的分割图像与原始图像相关性变化规律,建立了基于相关性比较的带宽自动计算方法;二是充分考虑核窗口内像素点颜色和位置的影响,提出以核窗宽内像素的中值作为收敛点的值,改善了图像分割的平滑性。在此基础上,针对静态图像中的多个头部目标的识别问题,在兼顾快速性和有效性的基础上,结合头部目标的发色分布特征和轮廓特征,建立了基于发色模型和圆存在性模型的级联检测器,以消除伪头部目标。实验结果表明,该方法能有效地消除与发色分布类似或灰度分布与头部接近的类圆区域,提高了头部目标识别的正确率。④研究了彩色图像序列中的多头部目标检测问题。以计算简单、快速、能较为准确地提取头部运动区域为目标,建立了一种结合运动检测与发色检测的多头部目标检测算法。该算法在研究和分析各种运动目标提取算法的基础上,提出基于3帧彩色边缘图像差分的运动目标区域提取,并通过对运动区域进行发色检测和连通域特征分析,实现对头部目标的快速提取和定位。⑤研究了基于改进mean shift的行人头部目标跟踪方法。在分析行人头部目标运动中的旋转、形变等特点的基础上,将mean shift算法引入行人头部目标跟踪,并针对mean shift跟踪算法的不足进行改进,给出了相应的解决方案。将LTP (Local Ternary Patterns)纹理特征引入头部目标跟踪,建立了融合LTP纹理特征和颜色特征的头部目标表示模型;提出了基于运动方向信息与核匹配的跟踪初始点选择算法,并采用基于主成分分析的跟踪窗口自适应跟踪技术。最终,结合上述工作,形成了一套复杂场景下的尺度自适应的行人目标跟踪方法。⑥基于前述研究成果,提出将行人头部目标的检测、跟踪与匹配技术有机结合起来,实现数目可变的多行人目标的自动检测与跟踪的方法,并将其应用于视频监控场景下的行人目标轨迹跟踪与公交场景下的上下车乘客人数自动统计。实验结果表明,所提方法能够解决新目标的出现、目标的暂时消失、目标误检测等问题,使得行人检测结果更加准确与可靠。各检测与跟踪算法实验以及综合的场景应用实验表明,本文提出的算法提高了单目视觉下行人目标检测的正确率,改善了行人目标跟踪的准确性,为解决复杂场景下多行人目标的检测与跟踪问题奠定了技术基础。更多还原

【Abstract】 Pedestrian detection is an important research area with many applications such as intelligent video surveillance, intelligent vehicle /driver assistance, motion analysis, and advanced human-machine interface, which has become the frontier and hot topic in the domain of computer vision in recent years. Pedestrian detection technology includes the pedestrian object detection, extraction, recognition, tracking and other aspects, and has some difficulties and challenges in occlusion, large variations of clothing, pose change, the arbitrariness and randomness of human movement, as well as the difference and complexity of application environment, which make it a difficult and challenging task and has received much attention from researchers.To avoid the occlusion among human body as much as possible, the pedestrian image of detection area was captured with vertical monocular camera and a method of detecting and tracking the multi-pedestrian based on head feature extraction was presented. Centering about the problem of multi-head detection and tracking under variation of formation and scale, the dissertation has made an in-depth study and discussion on head segmentation in static image, head identification based on multi-feature, fast detection of multi-head target in color image sequence, head tracking adapting to the change of object scale in complex scenes, thus the automatic detection and tracking methods of multi-pedestrian in image sequences was formed.The main study contents are summarized as follows:①The dissertation described the background, development and significance of pedestrian detection technology, based on the comprehensive comparison of various means of information perception, it focused on the status and research progress of computer vision-based pedestrian detection technology, and the drawbacks of current pedestrian detection technology were also discussed.②The dissertation made a detailed analysis of pedestrian information collection and the characteristics of pedestrian images with different collection way. Under the condition of serious occlusion among the human bodies, the full-body based and part-assembling based method was difficult to implement multi-pedestrian detection, and the head features were adopted to distinguish the multiple human bodies from complicated circumstances. Based on the analysis of the head features and the existed method of head detection under the vertical monocular vision, a novel method of head detection is presented by integrating multiple features such as color, contour and motion et al.③The dissertation studied on the segmentation and recognition technology of multi-head target in single color image.To solve the problem that the segmentation algorithm adopting the hair color or grayscale information was sensitivity to illumination and could not segment the head regions perfectly, an improved mean shift algorithm for color image segmentation was presented to extract head candidate regions. Compared with the conventional mean shift algorithm, the algorithm made improvements on two aspects: one is the adaptive bandwidth computing method on the ground of correlation comparison, which based on the similarity change relation between the original image and the segmented image with different bandwidth, the other is the better smoothness of the segmented image, which took the effect of pixels’color information and spatial information into account and took the median value of kernel window as the value of the convergence point.Furthermore, considering the trade-off between expeditiousness and effectiveness, the cascaded detector based on the contour information and inside color information of candidate head components was generated to implement the recognition of head. Experiment shows that the method can effectively eliminate fake regions whose color information is similar to hair color distribution or whose contour is quasi-circle, and the accuracy of head target identification is improved.④A method combination of motion segmentation and hair color partition was presented to implement multi-head target detection in color image sequence. On the basis of studying prior work, a algorithm based on three-frame-differencing and color edge information was proposed to perform the detection of moving object regions, and implement the extraction and location by using hair color segmentation and analyzing the connected component characteristics of candidates.⑤On the basis of analyzing the head motion characteristics of deformation and rotation, the mean shift method was introduced into pedestrian head tracking and gave out the improvement aiming at its drawback. Firstly, the algorithm took the LTP texture as the key feature for head tracking, and the target model was represented by fusioning LTP texture cue and color cue. Secondly, aiming at pedestrian tracking in large motion area, the selection method of initial point for mean shift tracking was proposed by combining motion direction information and kernel matching based on Bhattacharyya coefficients. Finally, the adaptive tracking window based on the principal components analysis was adopted and the pedestrian tracking method based on mean shift was formed, which could adapt to the change of object scale in complex scenes.⑥Based on the study mentioned above, the method integrating the technology of head detection, tracking and matching was proposed, and employed to implement the automatic multi-pedestrian detection and tracking with variable number, which was applied to the pedestrian trace tracking in video surveillance and passenger counting at bus entrance. The experiment show that the proposed method can solve the problem of the emergence of new object、temporal disappearance of object、false negative and false alarm of object, and make the detection result more accurate and reliable.The experiments of individual detection and tracking algorithms, as well as the integrated application in specific field indicate that the proposed algorithm improves the accuracy of pedestrian detection and tracking under the vertical monocular vision, and it lays the technology foundation for multi-pedestrian detection in complicated scenes.更多还原