【作者】 云廷进；

【导师】 郭永彩；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2008， 博士

【摘要】 与可见光成像相比,红外成像(特别是远红外成像)具有显著的优点。由于红外图像是热成像,可不依赖于外界光线条件。即使在黑暗和烟雾环境,也可获得在可见光波段无法成像的感兴趣的目标。由于人体目标是场景中最活跃和最有价值的要素,对场景中人体目标进行检测和跟踪一直以来是大家比较关心的问题。基于红外成像的人体目标检测和跟踪系统几乎可以在任何环境下进行全天候工作,而且在很多应用场合具有不可替代的作用,具有良好的应用前景。从技术角度,红外图像中的人体目标的检测和跟踪是一个具有挑战性的课题,内容涉及红外图像分割,人体目标的特征提取、特征描述和分类识别,非刚体复杂目标的运动跟踪等问题。要面对的主要困难有:(1)由于成像设备差异和环境因素变化造成的红外图像本身分辨率低,动态范围小,成像质量差等特点使得图像分割算法鲁棒性差;(2)由于红外图像是灰度图像,无色彩信息可用,并且成像模糊,纹理细节少,而人体目标是非刚体目标,姿态复杂多变,目标大小不一,因此如何有效地提取和描述红外图像中人体目标的特征,对目标和干扰进行分类识别困难很大;(3)人体目标的运动主观随意性强,无固定规律,并且目标的运动伴随着姿态和状态的改变,刚体目标的跟踪方法不适于人体目标的跟踪,同时由于红外人体目标用于跟踪的有效特征较少,可见光图像中一些优秀的基于颜色和纹理特征的人体目标跟踪方法不能直接使用。在对现有红外及可见光人体目标检测跟踪方法和红外图像中人体目标的成像特征进行分析后,本文提出了一套解决方案:(1)对于一般的红外图像,从红外成像的机理出发,提出了基于直方图多聚类分析的图像分割阈值选取方法,使用K-均值聚类中心分析法实现阈值快速选取;对于极性反转的红外图像采取传统的Mean Shift方法进行图像分割;对于发生互相遮挡粘连的人体目标,提出一种基于模板图像空间相对位置Mean Shift聚类分析的分割方法,后续的目标跟踪方法中采用了与这个算法类似的思想。(2)图像分割结束后根据得到目标的二值化模板特征选取候选目标。对因目标红外成像特征不一致而造成的目标模板缺损破碎现象,提出一种增强型的候选目标选取方法。初次候选目标选取完成后对剩余的模板图像使用图像距离变换,根据距离约束特征重新选取可能存在的目标区域并结合二次阈值选取算法,分割出潜在的候选目标。(3)对于候选目标,提出了基于网格划分的局部特征提取和特征描述方案。具体包括两个方法:局部梯度方向直方图描述方法和局部最大方向能量直方图描述方法。使用RBF—SVM分类器对候选目标进行分类识别,判断其是否为人体目标。(4)对人体目标的跟踪,提出了一种基于粒子Mean Shift迁移的跟踪方法,实现了对人体目标任意复杂运动形态的可靠跟踪。设计跟踪系统时采用面向对象的方法,对目标的特征进行封装,突出了跟踪问题本身。引入目标的有限状态机,结合目标检测算法,完成对新目标产生,目标隐藏、遮挡、消失等目标状况的处理。作者对方案各环节提出的算法进行了技术实现和实验验证,通过系统集成,设计了“红外人体目标检测系统”和“面向对象的红外人体目标跟踪系统”,在实际测试中取得了较为满意的结果。更多还原

【Abstract】 Infrared-image, especially the far-infrared image has prominent advantage comparing with the visible-image. As a result of thermal based imaging, infrared image is independent of external luminous qualification and is able to see the objects we interested in through darkness and frog, while the visible optical imaging is incapable to do this.Human detection and tracking are very important issues we concern with all along for humans are the most active and the most valuable factors in many occasions. Human detection and tracking system based on infrared imaging can almost work in any environment and all-weather conditions. It is irreplaceable in some situations and has vast potential applications in many aspects.However, this task is very challengeable from the technical viewpoint. It is involved with infrared image segmentation, human features extraction, description and classification, no-rigid complex objects tracking, etc. The main difficulties we have to confront firstly is the robust problem of the image segmentation algorithm which is caused by the drastic fluctuation of infrared image’s performance parameters because of the discrimination of different infrared imaging device , the environment changes and the intrinsic properties of infrared image, such as low resolution, narrow dynamic range and poor image quality. Secondly, because of infrared image is gray image, there are no color information available and some other drawbacks such as image blur, low textures, whereas human are non-rigid objects, their postures and appearances are complicated and changeable and their sizes are various, it is very difficult to extract and descript human features in infrared image effectively and to distinguish them from disturbances. Further more, human’s movements are very subjective and unbending; there are no routines to recapitulate them. Meanwhile, the movements also accompany with human postures and appearances transformation, the tracking methods used in rigid objects are not suitable for human tracking. One misfortune after another, being short of features for tracking, some excellent human tracking algorithms based on color information and textures can’t work efficiently for infrared human tracking.After studying on current human detection and tracking methods both in infrared images and visible images and analyzing on human imaging properties, the author presented an integrated solutions described as follows: (1) For general infrared images, we proposed a threshold selection method for image segmentation based on image histogram multi-clustering analysis proceeded from the infrared imaging mechanism. Specifically, we used K-means clustering centers algorithm to achieve the fast threshold calculation. For polarity-reversed infrared image segmentation we just adopted the traditional Mean Shift image segmentation algorithm. For occluded human segmentation, we proposed a relative position Mean Shift clustering algorithm in two-dimensional space applied to the object’s binary template, and we used the similar idea of this algorithm in subsequent human tracking system.(2) After image segmentation, we verified the valid candidates by the object’s binary template properties. For some half-baked and fragmented templates we proposed an enhanced candidates selection method. Carry out an image distance transform process on the surplus templates after the first candidates choosing process, then using the distance restriction to find the potential candidates corresponding regions. A second threshold calculation algorithm is invoked if necessary.(3) The feature extraction and description scheme of candidates are based on local area generated by grid division. There are two methods for doing this: Local Gradient-Orientation Histogram based feature description methods and Local Maximal Oriented Energy Histogram based feature description methods. We adopted RBF-SVM classifiers to predicate whether candidates were human or disturbance.(4) For human tracking, we proposed a tracking algorithm based on Particles Mean-Shift Migration algorithm. This algorithm is independent on any complicated movement manner of human. When designing the tracking system, we used the Object Oriented methods to encapsulating human features, the tracking problem’s substance was revealed. By introducing the finite-state machine technology and combining with the human detection method, the problems of new human coming, human hiding, occluding and disappear were handled properly.All the algorithms proposed in the scheme have been realized by the author and are validated by experimental results. The“Human Detection System for Infrared Images”and the“Object Oriented Human Tracking System in Infrared Videos”are designed based on the scheme by system integration; the performances of them are encouraging in practical application.更多还原