【作者】 胡威；

【导师】 蒋兴浩； 孙锬锋；

【作者基本信息】 上海交通大学 ， 电子与通信工程（专业学位）， 2013， 硕士

【摘要】 视频运动人体目标的跟踪技术已经成为了计算机视觉一个重要课题和研究热点，也有很大的现实意义和应用价值。本文研究对比了代表性的Meanshift、粒子滤波、Kalman滤波估计、光流法等视频目标跟踪技术算法，其中Meanshift算法是通过多次迭代漂移跟踪目标位置，对低速运动目标大多数情况都能达到实时、准确的跟踪，且简单易实现，但不适应快速运动目标或小目标的跟踪，存在误差累积错误跟踪。基于此，本文提出了一种结合Kalman滤波器和帧差法改进Meanshift算法的新跟踪方法。在Meanshift算法框架下，新方法依据Kalman滤波对目标质心和长宽预测，从帧差法检测到的运动区域中挑选可信区域只进行一次Meanshift颜色相似度检测，若不满足目标相似条件，则再对其他运动区域及Kalman预测区域依序进行原Meanshift算法跟踪，选出目标最佳跟踪区域。Kalman滤波器和帧差法可以预测和检测到丰富的运动区域信息，二者让Meanshift算法内核漂移跟踪更智能化。实验证明新方法减少了传统Meanshift颜色匹配迭代的次数和计算，在持久跟踪丢失目标的情况下也可以快速重新找回原目标准确跟踪，具有跟踪上的鲁棒性。最后本文拓展研究了人体目标跟踪技术的重要应用—视频人体动作识别，并提出了基于肢体的三个特征描述子了的动作识别方法。最后重申本文的研究创新点：1)在Meanshift算法框架下，设计一种结合Kalman、帧差法挖掘丰富运动信息新MeanShift跟踪算法，同样可以较好跟踪快速运动目标。2)新方法改善了原算法误差累积跟踪错误的现象，具有跟踪的自校正能力和鲁棒性。3)提出一种隔帧目标快速跟踪实验的设计。4)提出并设计了基于肢体的三个特征描述子的动作识别方法。更多还原

【Abstract】 The moving human tracking technology in video has become an important topic andresearch focus in computer vision.This paper has researhed some represented trackingmethods such as optical flow,Kalman Filter,Particle filter and Meanshift Algorithm.The ideaof Meanshift is to find the position of target through multiple iterations drift. It could be easilyrealized to get accurte tracking position for low-speed targets,but not suitable for tracking fastmoving target.There is also accumulated error to lost target in Tracking for Meanshift.In this paper,we proposed a new Meanshift tracking method combined with Kalman filterand frame difference method. Kalman filter could predict target’s centroid and scaling. Framedifference method can detect and get the information of movement areas. So the new methodguides Meanshift algorithm kernel tracking effectively in its drift process. Experiments showthe new method reduces the tracking time and is robust for long period tracking offast-moving targets.In the last chapter,we expand to research and design an action recognitionmethod based on three feature descriptors.Finally, we introduce the innovation of this study again:1) Design a new Meanshifttracking method with combination of Kalman filter and frame difference method.2) The newmethod is robust for fast-moving target in tracking with less time.3) Propose a frame-skippingtarget-tracking method.4) Propose and designe an action recognition method based on threefeatures of the limb descriptor.更多还原