【作者】 王勇；

【导师】 田金文； 谭毅华；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2009， 博士

【摘要】 随着计算机科学、电子技术、自动控制以及人工智能的发展与普及，智能视频监控技术已被广泛应用到国民经济的各领域，它在军事、工业、智能人机交互、智能交通和科学研究等方面都具有重要的意义，应用前景广阔。作为智能视频监控的核心关键技术，序列图像目标检测与跟踪在理论和应用上仍存在许多问题和难点尚未解决。很多因素都会影响在序列图像中对运动目标的可靠观测，例如，视觉特征分辨力较弱、背景干扰、目标间相互遮挡，加上实际环境中目标运动的随机性和复杂性(目标大小、形变、运动速度、光照变化、目标颜色与背景颜色的相似程度等)，这些都使得算法的设计变得非常困难。因此，研究复杂场景下的目标检测与跟踪具有重要的理论意义和很高的实用价值。论文针对复杂场景中运动目标检测与跟踪中存在的若干问题，对运动目标的准确分割、区域特征提取、目标描述及鲁棒跟踪等关键技术进行了研究与探讨。论文的主要研究内容和成果概括如下：(1)对静态背景下运动目标的检测与分割进行了研究。针对复杂背景、光照变化、阴影等影响目标检测的问题，提出了一种有效的自适应混合高斯背景建模算法，各像素点根据其像素值出现的混乱程度采取不同个数的高斯分布描述，重新设计了背景模型的学习与更新方法以及高斯分布生成准则；采用基于联合像素时空信息的分割与形态学重构方法使得前景目标分割的性能得到了有效地提高。同时文中也给出了阴影检测与抑制、光照突变的处理方法。该算法能够快速准确地建立背景模型，准确分割前景目标。(2)针对固定监控场景提出了一种基于色彩分割与目标局部模型匹配的非刚性目标跟踪算法。利用自适应混合高斯背景模型提取前景运动目标，通过基于均值漂移滤波与区域生长的色彩分割算法建立目标局部模型并实时更新，结合区域约束条件和目标模型区域匹配实现目标的跟踪。算法能有效地解决跟踪过程中目标局部遮挡与形变问题。(3)针对基于核的跟踪算法中颜色直方图对目标特征描述较弱、跟踪过程中核函数带宽保持不变、无法实现模板更新的缺点，提出了一种基于偏移校正的核空间直方图目标跟踪算法。在特定的色彩空间中，联合像素颜色与空间信息建立目标模型，即建立目标颜色直方图的同时用高斯分布对直方图各区间内的像素坐标建模，有效地解决了目标与背景颜色相似情况下的跟踪问题。通过目标特征点匹配估算目标仿射模型参数实现跟踪窗的更新。结合卡尔曼滤波器较好地克服了均值漂移算法不能跟踪快速目标的问题，计算卡尔曼残差来判断目标是否发生遮挡，从而选择卡尔曼滤波或是线性预测来估计目标位置。通过对跟踪过程中目标模板的偏移校正来消除跟踪误差的累积，保证目标空间直方图能准确描述目标，提高跟踪的鲁棒性。(4)为了解决单一视觉信息在环境变化，以及目标平移、旋转和尺度变化等情况下描述目标不够充分、目标跟踪不够稳定的问题，提出了一种基于自适应混合似然模型的粒子滤波跟踪算法。在状态矢量中增加表示目标几何变换的变量，通过仿射变换来处理目标的运动变化。目标观测的描述融合了多种特征信息，似测模型能根据当前的跟踪环境的变换实时切换。由于算法始终利用对当前跟踪场景可分性好的特征信息跟踪目标，有效地提高了粒子滤波跟踪算法在复杂场景下的稳健性。论文对复杂场景下运动目标检测与跟踪算法进行了比较深入的研究，重点针对目标分割、目标建模、遮挡、鲁棒跟踪等关键技术问题，提出了新的、有效的解决思路和实现方案。更多还原

【Abstract】 With the development and popularization of computer science, electronic technology,automatic control and artificial intelligence, the intelligent visual surveillance plays veryimportant roles in military, industry, human-computer interaction, intelligent transformand science researching, etc, and it is of a wide development prospect as well. So peoplepay more and more attention to the researches of sequential images motion detection andvisual tracking, which is a key technology of the intelligent visual surveillance and is anactive research topic in the image processing and computer vision. At present, movingtarget detection and tracking is not well-considered, many problems and difficult points intheory research and in applications are still unsolved. However, many issues such as weakdistinguishing image features, background clutter, occlusion, in addition, the targetmovement often behaves very complicated in real environment (e.g. target size, shapechanging, move speed and path, illumination changing, the similarity of target color andbackground color, and background stability, etc), can affect the effective observation of thetracked targets in images and make robust tracking algorithms designing a very difficultproblem. Therefore, the researching on moving target detection and tracking undercomplicated background has both important theory significance and application value.Aiming at resolving the difficult problems of the robust visual tracking, we studied severalkey technologies under complicated environment from movement target detection andsegmentation, target area feature acquiring, target describing, and robust tracking. Themain contents and contributions of this dissertation are summarized as follows:(1) The moving target detection and segmentation have been studied in stationaryscene. An effective adaptive background updating method based on Gaussian mixturemodel (GMM) was presented. The number of mixture components of GMM is estimatedaccording to the frequency of pixel value changes, the performance of GMM can beeffectively improved with the modified background learning and update, new distributiongeneration rule and morphological reconstruction based on spatial and temporal pixelsinformation. The detection of illumination great change and shadow removal were alsoproposed.(2) To enhance the performances of object tracking in stationary scene, a trackingmethod based on adaptive color segmentation and object part model was presented. In thiswork, the foreground blobs are obtained by background subtract. Object parts in the part model are generated online by the color segmentation based on mean shift andregion-growth. The constraints between parts and region features are taken into accountand used to perform objects tracking. The algorithm can solve the partial occlusion andobject deformation problem well.(3) To improve theoretic limitation of the traditional Mean shift, a novel targettracking algorithm was presented. Firstly, a new color space is partitioned into subspacesby considering the weighted number of pixels with feature vectors cluster, and describingthe pixel coordinates with Gaussian distribution. Then the target model and the candidateare constructed through an improved spatial histogram, which has ability to surmount thesimilarity between target and background. Finally, the affine transform is establishedcombining comer and edge detector to update tracking window. The algorithm is able tohandle the fast target tracking and occlusion by combining kalman filter and Mean shift.Besides, the algorithm has better effect and robust through eliminating the tracking errorusing the drift correction of target template.(4) To resolve the problem of single visual cue in different enviroments and posechanging, a particle filter tracking algorithm based on cues integration mechanism andadaptive observation likelihood is proposed. The multiple cues are adopted to representtarget, the likelihood model is constructed on-line with the reliable cues. Besides, hiddenvariables indicating geometric transformation are also augmented in the staterepresentation, and affine transformation is used to handle the movement changing oftarget.In the dissertation, the target detection and tracking algorithms have been studiedextensively under complicated environments, robust and practical methods were proposedto solve the key technologies, such as target segmentation, target modeling, occlulsion,and robust target tracking.更多还原