CASIA OpenIR  > 毕业生  > 博士学位论文
基于视觉的动态场景下目标跟踪方法研究
其他题名Research on Object Tracking in Dynamic Scenes Based on Computer Vision
尹春霞
学位类型工学博士
导师徐德 ; 李成荣
2012-05-30
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业控制理论与控制工程
关键词目标跟踪 动态场景 Mean-shift 粒子滤波 特征检测 Object Tracking Dynamic Scenes Mean-shift Particle Feature Detection
摘要目标跟踪在科学和工程中具有重要的研究价值。由于摄像机和目标之间的相对运动以及应用场景的复杂多变,采集的视频图像中一般具有光照变化明显、图像中杂物或噪声显著、目标被部分遮挡或完全遮挡、目标姿态变化大等特点。针对上述问题,本文结合均值漂移、粒子滤波、特征点检测以及视觉注意机制等知识,对动态场景下的目标跟踪方法展开深入研究。本文的主要内容和贡献如下:  分析了常规均值漂移(Mean-shift)跟踪算法与带宽自适应均值漂移跟踪算法的优点及不足,讨论了跟踪目标初始化过程中的目标检测方法,提出使用基于游程链的Blob全局搜索算法对丢失的目标进行全局搜索。  提出了一种改进后的带宽自适应均值漂移算法,并给出了实现步骤,进行了实验验证。改进算法继承了自适应带宽均值漂移算法的优点,能够解决带宽选择问题和带宽更新问题,实现了跟踪目标初始化与目标丢失后的全局检测。  针对常规算法对光照变化、局部遮挡和临时性完全遮挡问题不具备鲁棒性,易陷入局部最优状态的问题,本文以均值漂移算法为基本框架,使用数据融合策略判断目标状态,提出了融合粒子滤波的ABMSPF(Adaptive Bandwidth Mean-shift Method Assisted by Particle Filter,即融合粒子滤波的自适应带宽均值漂移)跟踪方法;同时提出一种用于快速搜索目标的局部搜索策略。对ABMSPF方法、改进后的带宽自适应均值漂移方法和粒子滤波跟踪方法进行实验对比研究,从复杂场景下的跟踪稳定性、目标定位精度、计算复杂度等几个方面进行分析比较。实验表明ABMSPF方法在光照变化、部分遮挡、短时间内的完全遮挡、相似色彩干扰等复杂场景中均能准确地跟踪目标,并且跟踪轨迹平滑。  针对常用的Harris,SIFT等局部特征检测方法,从算法原理和实验数据两个方面进行分析,比较了在旋转、尺度、光照、视角等参数变化时,上述算法的鲁棒性、稳定性和计算复杂度。  使用特征点匹配的方法检测目标时,特征点数量大、运算时间长。针对这一问题,考虑将视觉注意机制应用到图像的特征提取与特征匹配过程中,提出了一种基于显著图的SIFT特征检测与匹配方法,并探讨了该方法在目标跟踪中的应用。
其他摘要Object Tracking based on computer vision is important in scientific research and project. For the existence of relative motion between object and the camera, or the complexity of scenes, there may be many problems in video sequences such as changing illumination, serious debris or noise, partial or complete occlusions, or the object state may be changing from time to time. All of that make object tracking in various scenes difficult to be achieved. In view of these problems, this paper makes use of knowledge in mean-shift, particle filter, feature detection as well as the visual attention mechanism, to study the object tracking methods under dynamic scenes. The main contents and contributions of this paper are as follows:  The advantages and shortcomings of conventional Mean-shift tracking algorithm and the bandwidth-adaptive Mean-shift algorithm are analyzed, and the object initialization methods are discussed. A run-list based Blob searching method is proposed to search the target globally once it is lost.  An improved bandwidth-adaptive algorithm is proposed, then implementation steps are provided and at last experiments are carried out to check the property of this algorithm. The improved algorithm has advantages of bandwidth adaptive algorithm. It updates the object bandwidth through iteration, realizes automatic initialization of the object and searching of the missing object.  None the above algorithm is robust to illumination changes, similar interference or occlusions. In view of that, a tracking algorithm called ABMSPF (Adaptive Bandwidth Mean-shift Method Assisted by Particle Filter) is proposed. This algorithm takes Mean-shift algorithm as a framework, combining with particle filter which is used to provide auxiliary positioning information, and the target state is determined by a data fusion strategy. Properties such as stability, location accuracy and calculating complexity of ABMSPF, improved bandwidth-adaptive Mean-shift and the particle filter methods are analyzed experimentally. The ABMSPF algorithm is proved to be able to track object accurately in complex scenes with illumination changing, occlusions or color interferences, and the tracking trajectory is smooth.  Under situations such as image rotating, scale changing, illumination, or perspective changing, several commonly used feature detecting methods like Harris and SIFT are analyzed, the properties of robustness, stability and computing c...
馆藏号XWLW1724
其他标识符200818014628026
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/6450
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
尹春霞. 基于视觉的动态场景下目标跟踪方法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2012.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA_20081801462802(2733KB) 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[尹春霞]的文章
百度学术
百度学术中相似的文章
[尹春霞]的文章
必应学术
必应学术中相似的文章
[尹春霞]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。