【作者】 孙宸；

【导师】 周越；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2012， 硕士

【摘要】 目标跟踪,属于计算机视觉的研究范畴,作为计算机视觉领域中一个极具吸引力又极富挑战性的课题,一直得到研究工作者们的广泛关注和投入。利用机器学习方法解决跟踪问题是最近兴起的热点研究领域,也代表了未来目标跟踪领域的方向。本文对基于限制条件的半监督学习算法在目标跟踪中的应用进行了研究,主要工作如下:1.在分析与对比现有的半监督学习算法的基础上,对P-N Learning进行了深入研究,并通过实验验证了P-N Learning的性能。2.对多种跟踪算法进行了研究。主要包括经典的光流法,新近流行的仿射粒子滤波算法以广义霍夫变换在目标跟踪中的应用。将光流法和仿射粒子滤波分别纳入本文设计的半监督学习跟踪系统框架。在广义霍夫变换的研究方面,实现了一种基于广义霍夫变换的融合局部特征匹配与概率外观模型的多目标跟踪方法。3.设计实现了一种基于半监督在线学习的跟踪系统框架。利用多层限制条件对训练样本选取和分类器训练进行干预,提升了分类器性能。并分别将光流法和仿射粒子滤波纳入系统框架,实现了完整的跟踪功能。实验证明,本文系统在目标快速运动、目标旋转、目标消失重现、外观相似目标遮挡等情况下,都能取得较为理想的跟踪效果。更多还原

【Abstract】 Object tracking is an attractive and challenging subject of computer vision, draws a lot of attention. Solving tracking problem with machine learning technique is a new rising area of research, also indicates the trend of tracking research in the future. In this thesis, we mainly study the application of semi-supervised learning in object tracking area, especially the one based on structural constraints. It’s organized as follows:1. Semi-supervised learning algorithm Several popular semi-supervised learning algorithms are studied, including semi-boosting, multiple instance learning, and P-N learning. Main attention is paid to P-N learning, especially its performance in object detection and tracking area.2. Research on tracking algorithm We investigate several popular tracking algorithms, including optical flow, particle filter on affine group, and generalized hough transform. Optical flow and particle filter on affine group are incorporated in our semi-supervised learning based tracking framework. Moreover, A new approach for tracking multiple objects is proposed, which combines feature correspondence with a probabilistic appearance model, and uses generalized hough transform to determine the optimal target position.3. Design and implementation of tracking system based on semi-supervised learningTemporal, spatial and data correlation constraints are designed to intervene extraction of training data, training of classifiers and labeling of unlabeled data, which improves the performance of classifiers. Optical flow and particle filter on affine group are incorporated in our tracking system framework. Experiments show that our tracking system work well under challenging situation, including fast moving, large scale rotation, object reappear and occlusion with similar object.更多还原