导　　师： 丘水生

学科专业： 080902

授予学位： 博士

作　　者： ;

机构地区： 华南理工大学

摘　　要： 基于视频图像的目标跟踪是计算机视觉的核心问题,相关研究在人机交互、智能视频监控、智能交通等领域有广泛应用前景。跟踪算法的核心部分是目标外观建模,由于目标在运动过程中形态往往不可预知,使得跟踪算法研究中具有在线学习能力的外观建模成为热点。近年来,尽管相关研究取得了众多进展,但在实际应用中对任意目标实现可靠跟踪仍是难点,主要受到目标形变的复杂程度、场景中相似干扰、无法判断跟踪可靠性等问题影响。目前,研究者们从目标外观模型的特征设计、目标表示、模型更新方法以及整体跟踪框架中多种信息的融合等方面开展深入研究,试图解决上述跟踪难题。本文属于这一类研究。 本文主要做了两方面的工作：1/)以场景中相似干扰多导致跟踪困难的自然交互“设备”手部为研究对象,从考虑目标特点的外观建模,背景信息、目标运动信息的综合运用等角度,研究了复杂场景中鲁棒的手势跟踪方法；2/)以特征信息丰富但因形变复杂造成跟踪困难的目标为研究对象,模拟人眼初级视皮层腹侧通路的形状感知机制,研究了鲁棒的目标外观建模方法。主要研究内容包括： 1.对近年来文献中出现的目标在线跟踪方法进行了比较研究。将六种具有代表性的目标跟踪方法进行了实现,选择涵盖目标形变复杂、光照变化剧烈、存在相似干扰、遮挡严重等特点的测试视频,从定性的跟踪效果与定量的实验数据进行算法的对比分析。在此基础上,总结出了现有方法对本文研究对象进行跟踪时存在的不足,并指出了可能解决问题的研究方向。 2.以手部为研究对象,提出了多信息融合的增量子空间手势跟踪方法：1/)结合手部的肤色聚类特点,将增量子空间学习算法扩展应用到彩色图像,提出了融合亮度信息与颜色信息的彩色增强型增量� Object tracking is an important issuein the field of computer vision, and is widely used in human-computer interaction, intelligent video monitoring, intelligent transportation and so forth. Appearance model is thecore part of object tracking framework, among which the ability of online learning is becoming a hot research area because the object appearances are always unpredictable. Although great progress has been made in the research of object tracking in recent years, it is still difficult to track an arbitrary object in real environments, due to influencesby complex deformations of the object, similar interferencesin the scene, unknowns of tracking reliability and so on. Researchers are trying to solve these problems from feature design of object's appearance, representation of object, updating mechanism ofappearance model and fusion ofmulti-cues in the tracking framework. This dissertation belongs to this kind of research. This dissertation focuses on two parts of work:1/) Hand, as an important interaction 'device', is treated as the tracking object in the research, which is difficult to be tracked because oftoomany similar disturbances in the scene. The robust hand tracking algorithm in complex environment is studied from the appearance modeling by consideration ofobject's characteristics, and comprehensive utilization ofbackground and motion information.2/) Objects with rich features are still hard to be tracked because of its complex deformation. For objects belonging to this kind, the robust object appearance modeling algorithm is studied by imitating the shape perceptive mechanism in the ventral way of human primary visual cortex. The main contributions of the dissertation are summarized as follows: 1. Online object tracking algorithms appeared in recent literature are studied and compared. Six typicalalgorithms areimplemented and explored. The testing videos areselected, containingcomplex deformation of the target, illumination with dramatic changes, similar disturbances in the scene and severe

关 键 词： 目标跟踪 外观模型 增量子空间学习 光流 生物启发特征

领　　域： [自动化与计算机技术] [自动化与计算机技术]