【作者】 权伟；

【导师】 陈锦雄；

【作者基本信息】 西南交通大学 ， 信号与信息处理， 2013， 博士

【摘要】 对象跟踪作为智能视频分析的关键问题,在计算机视觉领域具有广泛的应用,如智能监控、人机交互、机器人技术以及多媒体应用等。尽管研究人员对此做出了大量的工作,但是由于真实世界的复杂性,如背景干扰、表观变化、低图像分辨率,以及帧跳跃等问题,使得在无约束环境下实现对目标进行长时间实时稳定的可视跟踪任务仍然是一项极具挑战性的课题。本文通过对对象跟踪算法研究现状的分析,结合跟踪过程具有很强的时序性和时空关系的特点,基于图像信号分析、模式识别和在线机器学习的理论与方法,分别从单目标跟踪、多目标跟踪及其应用三个方面开展研究,提出了几种实时稳定的对象跟踪算法。具体工作如下：(1)为了提高依赖随机蕨检测的跟踪算法的稳定性,提出了一种基于增强型随机蕨的对象跟踪算法。该算法在学习过程中,通过在线聚类随机蕨每个叶节点中的学习样例,自动发掘其特征空间中特征向量潜在的分布特性,即隐含类型；在评价过程中,将这些隐含类型作为核函数的数据点进行核密度估计,计算测试样例的类型概率。实验结果表明,该算法在实现实时对象跟踪的同时提高了跟踪的稳定性。(2)针对基于在线学习的跟踪算法面临的两难问题,即如何既保证对目标变化的适应能力,又保证学习的准确性,提出了基于主动场景学习的对象跟踪算法。该算法基于对象与背景信息建立结构化的约束,并根据该约束对在线模型和检测器进行有监督的学习,从而提高了其学习的准确性。同时结合基于光流分析的目标运动区域提取方法,使得能够对快速移动目标进行跟踪。实验结果表明,该算法提高了跟踪系统对目标变化的适应能力和跟踪的稳定性。(3)针对基于霍夫变换的对象跟踪算法难以实现实时跟踪的问题,提出了基于霍夫蕨的对象跟踪算法。该算法采用依赖检测的跟踪框架,以随机蕨作为基础检测结构,将对象的局部表观作为学习数据,在其每个叶节点中计算并保存霍夫空间中属于目标对象的投票概率,并通过在线学习使其能够同步适应对象表观的变化。实验结果表明,该算法在满足跟踪稳定性的同时能够实现实时的对象跟踪。(4)为了提高跟踪过程中检测器的对象识别能力进而提高跟踪的稳定性,提出了在线学习多重检测的对象跟踪算法。该算法将目标对象的整体和局部表观,以及由场景学习中发掘的同步对象同时作为学习数据,因此能够在跟踪过程中分别对这些类型的对象进行检测。最后通过计算这些检测结果关于目标的配置概率进而确定目标的位置,实现对象跟踪任务。实验结果表明该算法可以适应更加复杂的跟踪环境,在满足实时性的同时提高了跟踪的稳定性。(5)为了降低多目标跟踪算法的计算复杂度,实现实时的多目标跟踪,提出了基于自适应运动相关协作的多目标跟踪算法。该算法根据目标运动信息建立目标间相关度,通过相关度状态估计协作模型预测目标状态,实现多目标跟踪。实验结果表明,仅采用基本的短时跟踪算法,结合该协作模型则可以有效的处理目标遮挡,实现实时稳定的多目标跟踪。(6)针对对象跟踪算法的应用问题,结合具体的应用场景研究了相应的技术方法。关于医学图像处理的应用,提出了一种基于分层检测的人体膝关节前交叉韧带(ACL)定位方法,用于解决在图像中检测和定位前交叉韧带区域问题,从而促进前交叉韧带重建手术的研究。该方法将韧带定位分为全局与局部检测,根据不同的样例图像选择不同的图像特征,基于随机森林构建对应的全局和局部检测器,通过确定膝关节中前交叉韧带的整体组织的位置,再进一步识别属于前交叉韧带的具体区域,从而实现对它的准确定位。基于真实人体膝关节MRI图像的实验结果表明,该方法对前交叉韧带的检测识别能力高,且定位准确。更多还原

【Abstract】 This thesis mainly focuses on the problem of visual object tracking, which is a key problem of intelligent video analysis that is demanded by many applications in computer vision, such as intelligent surveillance, human-computer interfaces, robotics and multimedia. Robust long-term visual tracking in unconstrained environment is still very challenging due to the real-world complications such as clutters, appearance change, low image quality, and frame-cut. Based on the analysis of research actuality of object tracking which contains strong spatial-temporal relevance and the theory and method of image signal processing, pattern recognition and online machine learning, we propose several robust real-time object tracking algorithms, involving single target tracking and multiple target tracking, and apply them to address other problems in computer vision. The main contributions of this thesis are given as follows:(1) In order to improve the robustness of tracking algorithm using random ferns for detection, we propose an enhanced random ferns which is integrated into our tracking framework as the object detector. Its main idea is to exploit the potential distribution properties of feature vectors which are here called hidden classes by on-line clustering of feature space for each leaf-node of ferns. The kernel density estimation technique is then used to evaluate unlabeled samples based on the hidden classes which are set as the data points of the kernel function. Experimental results demonstrate the effectiveness and the improved robustness of our approach.(2) To address the problem of improving the ability of adaptation to the variation of target and meanwhile ensuring the accuracy of online learning for tracking system, we propose a method of active context learning for object tracking. The approach exploits both target and background information on the fly automatically and builds the structural constraint by using the active context learning to enhance the adaptability for variation of the target and stability of tracking. An optical-flow-based motion region extraction method is integrated into the context learning framework to address the problem of fast target motion or abrupt camera motion. Experimental results demonstrate the improved tracking performance of our tracker.(3) Existing Hough-based tracking systems have not achieved real-time performance. To deal with this problem, we propose a Hough ferns based method for real-time object tracking. In the tracking-by-detection framework, Hough ferns, which are based on random ferns, sample the local appearances of object as training set, and compute and save the Hough votes for each leaf-node. Hough ferns and object model are leaned on-line to adapt to the variation of object. Experimental results validate the effectiveness and robustness of our tracker which can run in real time.(4) In order to improve the capability of object recognition of the detector and then the robustness of the tracking system, we propose a method of online learning multiple detectors for object tracking. The method uses the random ferns as the basic detector. The entire and the local appearances of the target and the connected objects which are explored by the context learning are used synchronously as the training data to build and upgrade the object detector on-line. Thus it is able to detect the objects with different classes independently. Since different detection is related to different object class, the results of object detections are fused as the measurements and the probabilities of configuration hypotheses for the measurements to the target are calculated to find the target location for visual tracking task. Experimental results validate the effectiveness and robustness of our approach and demonstrate its better tracking performance than several state-of-the-art methods.(5) To reduce the computational complexity of the algorithm achieving real-time multiple target tracking, we propose a collaboration model in which the acceleration difference between two targets is used to calculate the motion correlation value based on the two-dimensional Gaussian function. By the collaboration model, the location of occluded target is estimated using the motion information from other targets. The proposed approach is computationally efficient and robust. Experimental results exhibit the performance of our tracker based on our approach.(6) For the application of object tracking, the methods proposed can be applied to the corresponding scenarios. In particular, in order to address the problem of detecting and locating the anterior cruciate ligament of human’s knee in medical image and promote the study of its reconstruction operation, we proposes a hierarchical detection based method to locate the anterior cruciate ligament. The location task is considered to be to perform the global and the local detections successively. The features are selected according to the type of image samples, and the corresponding global and local detectors are built based on the random forests respectively to first find the entire region of the anterior cruciate ligament and then recognize its definite area. Experimental results based on the real MRI images validate the effectiveness and accuracy of our method.更多还原