【作者】 邱雪娜；

【导师】 刘士荣；

【作者基本信息】 华东理工大学 ， 控制理论与控制工程， 2011， 博士

【摘要】 运动目标跟踪一直是计算机视觉领域备受关注的前沿课题之一,在移动机器人定位与导航、多机器人编队、月球探测以及智能监控等方面都有非常重要的应用。经过几十年来学者们的不懈努力,运动目标检测与跟踪技术取得了长足的进步。但是,由于视觉跟踪系统应用环境复杂性(比如光照、遮挡等因素影响)以及目标本身的多样性,给目标检测、跟踪技术带来了极大的困难。本文在研究传统视觉目标跟踪算法的基础上,对单目标跟踪、多目标跟踪、广义目标检测、立体视觉跟踪等问题,结合学术前沿知识,提出了新的算法,提高跟踪的准确性和鲁棒性。最后还应用到移动机器人目标识别、定位与跟踪上,具有较好的跟踪性能。本文研究工作可以总结为以下几个方面：(1)为了克服单个目标跟踪算法在复杂环境下跟踪精度不高的问题,提出了一种基于序贯检测机制的运动目标跟踪算法。该算法在序贯检测机制下,将粒子滤波、稀疏场主动轮廓和Camshift等算法结合。首先用基于颜色特征的粒子滤波估计最优跟踪窗口;而此跟踪窗口和目标的相似度决定是否采用稀疏场主动轮廓算法,同样目标轮廓和目标的相似度决定是否需要Camshift对轮廓进行修正。跟踪实验表明本章所提的算法在不同的复杂环境和目标有尺度、旋转、视角等姿态变化的情况下,具有较好的跟踪精度和鲁棒性。(2)针对单一视觉特征的目标模型很难对环境中所有变化都具有足够的鲁棒性,提出了一种自适应多特征融合的核函数目标跟踪算法。该算法目标模型的概率分布用SIFT、颜色和运动特征的核函数线性加权来表示。目标的表征特征和运动特征相结合,可以提高跟踪的稳定性和精确性。跟踪窗口尺寸根据匹配到的SIFT特征对的仿射变化参数实时调整。每个特征核函数的权重跟随特征的显著性变化而自适应变化,从而可以最大限度地发挥每个特征的作用。实验表明该算法可以在不同的场景下跟踪目标,而且可以应对目标有姿态、尺度、旋转、视点变化和环境光照变化等情况,而且本算法的跟踪性能优于Camshift算法、基于SIFT特征目标跟踪、基于彩色SIFT目标跟踪算法。(3)针对多目标跟踪问题,提出了一种基于信息分享机制的粒子滤波多目标跟踪算法,提高多目标跟踪效率。该算法将粒子群优化算法和蚁群优化算法的优化思想共同作用到粒子更新中,实现粒子之间信息共享,从而增强粒子的多样性和最优估计能力；同时分析了该算法的收敛性。实验表明,本算法能用较少的粒子达到较高的跟踪精度,而且跟踪性能优于传统粒子滤波和基于粒子群优化的粒子滤波算法。(4)研究了广义目标检测算法,针对广义目标的复杂性和多样性,提出了一种嵌入Bag-of-words的Boosting目标检测算法。Boosting算法具有较好的检测效率,但是对于广义目标的复杂性和多样性,会存在一定的误检测,本文算法嵌入Bag-of-words算法,用其基于局部块特征,而且简单、鲁棒性好,对遮挡、复杂目标具有较好的分类性能等特性,对Boosting检测结果进行修正,剔除其误检测部分,从而提高广义目标检测精度。(5)针对双目视觉比单目视觉具有更丰富的信息量,研究了基于双目视觉的目标跟踪算法,提出了两种基于双目视觉的移动机器人实时动态目标识别与定位的算法。一种算法首先采用SIFT算法提取目标特征,并结合双目视差特征进行目标匹配；然后通过区域增长算法进行目标区域的提取；最后结合双目视觉标定模型对目标进行定位。另外一种算法把基于序贯检测机制算法、双目视觉视差信息、标定模型相结合来完成双目视觉跟踪,再引入视差置信区间判据可有效减少噪声影响,提高运动目标定位精度。实验表明这两种算法在摄像机运动-目标运动情况下,能对动态目标进行有效地识别与定位。更多还原

【Abstract】 Dynamic target tracking is always given to attention for computer vision, and it is widely applied in robot location and navigation, multi-robot formation, lunar exploration and intelligent surveillance. Significant progress has been made in targets tracking during the last few years. However, the task of robust tracking is challenging due to the environment complexity (such as fast illumination variation, occlusion) and object diversity.On the basis of traditional object tracking methods, the thesis aims to single object tracking, multi-object tracking, generic object detection and stereo vision object tracking. It combines academic frontiers and presents new algorithms to improve the accuracy and robustness of tracking performance. Moreover, the proposed method is applied to mobile robot object identification, localization and tracking, and it has better tracking performance.The main research work of this dissertation can be summarized as follows:(1) A novel object tracking method is presented based on sequential detection mechanism to improve tracking accuracy of single object tracking method in a complex environment. The proposed method integrates particle filter, sparse field active contours and Camshift under sequential detection mechanism. First, particle filter object tracking is applied using color feature to estimate optimal tracking window and then sparse field active contours is performed based on the similarity of object and resulting windows. Similarity detection is carried out again to determine whether Camshift should be employed to modify the object contour and thus achieve accurate tracking of moving object. Experiments demonstrate that the proposed method based on sequential detection mechanism can effectively track and locate the moving object, and it can handle both target changes in scale, orientation, view and environment illumination changes, moreover, it is able to track and locate the object accurately.(2) A new adaptive kernel-based target tracking method with multiple features fusion is proposed to solve object model based on single visual feature not having enough robustness when environment changing much, A linear weighted combination of three kernel functions of scale-invariant feature transform (SIFT), color and motion features is applied to represent the probability distribution of the tracked target. Appearance and motion features are combined to enhance the target region location stability and accuracy. The size of the tracking window can be adjusted in real time according to the affine transform parameters of the corresponding SIFT couples. The weights of three kernel functions are also adaptively turned according to the scene, in order to extremely excert the function of the features. Experiments demonstrated that the proposed algorithm can track the moving target successfully in different scenarios. Moreover, it can handle target pose, scale, orientation, view and illumination changes, and its performance is better than the classic Camshift algorithm, SIFT tracking method and color SIFT tracking method.(3) A novel particle filter algorithm based on information shared mechanism is proposed for multiple object tracking and it can improve tracking efficiency. The proposed method combines particle swarm optimization and ant colony optimization to update particles, and thus population information is fully shared. As a result, it can recover particles diversity and increase the precision of the estimation. Moreover, the convergence of this algorithm is analyzed. The results of visual tracking experiment show that the presented algorithm can realize multi-object tracking with fewer particles and its combination tracking performance is better than classic particle filter and the particle filter based on particle swarm optimization, which demonstrates the effectiveness of the proposed algorithm.(4) It studies generic object detection. In the view of complexity and diversity of generic object, it proposes Boosting generic object detection method with bag-of-words. Boosting method has good detection efficiency, but it has some fault detections due to the diversity and complexity of object. Bag-of-words method has some advantages, such as local patch features, simplicity and robustness, and it has good classification performance of complex object. The proposed method applies bag-of-words to remove the fault detection and to improve the tracking results of Boosting, and thus it achieves high generic object detection accuracy.(5) For binocular vision hasing abundant information than monocular vision, this thesis focus on binocular vision tracking and it presents two real-time dynamic object recognition and localization methods for mobile robot using binocular vision. For one, firstly, the SIFT operator is applied to object features extraction and object matching with the disparity features of binocular vision. Then the object area is extracted through region growing method. Finally, according to the binocular vision calibration model, the object’s location is obtained. For another, it is combined sequential detection scheme, disparity information and binocular vision calibration model to accomplish binocular vision tracking. The disparity confidence interval criterion is introduced to decrease the effect of the noise effectively and enhance the object localization accuracy. Experiments demonstrate that the proposed methods can effectively recognize and locate the dynamic object in both camera moving and object moving.更多还原