【作者】 林高鹏；

【导师】 张茂军；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2009， 硕士

【摘要】 全向视频是虚拟现实和计算机视觉中一种重要的场景表示方法,它指的是在垂直方向180°和水平方向360°的图像视图。与传统视频相比,全向视频的优点在于其360°的宽广视野感知范围,全向摄像机的运动估计比传统摄像机更加稳定,因为传统摄像机视野小,运动估计对摄像机的方向很敏感,而且其水平方向上平移产生的运动与绕垂直轴旋转产生的运动十分相似,难以区分,而全景摄像机特有的360°视野能避免这些问题。正是由于具有这些优势,基于全向视频的目标检测与跟踪已经在一些领域开始得到越来越多的研究,例如视频监控、机器人导航和智能驾驶等领域。本文在分析全向图像特点的基础上,针对经典目标检测与跟踪算法运用于全向视频的不足,分别对目标检测算法和目标跟踪算法加以改进,论文的研究成果主要包括以下几个方面:(1)在分析马尔可夫随机场图像运动检测原理的基础上,根据全向图像的分辨率和邻域特点,提出一种新的折反射全向图像邻域定义方法。(2)在分析卡尔曼滤波器图像运动目标跟踪原理的基础上,针对目标沿直线运动时在折反射全向图像上的成像轨迹是二次曲线的问题,修改卡尔曼滤波器,提高折反射全向视频卡尔曼滤波目标跟踪时的状态预测精度。结合以上方法,通过多个实验视频来验证算法的鲁棒性和准确性,其实验结果令人满意。更多还原

【Abstract】 Omni-directional video is an important representation for scenes of virtual reality and computer vision, which refers to the image of vertical 180°and horizontal 360°view. Compared with traditional video, the advantages of the omni-directional video lies in its 360°field of vision sensing range. The motion estimation based on omni-directional camera is more stable than traditional cameras. The traditional cameras has a small vision, so motion estimation is very sensitive to the direction of the camera, and its movement of horizontal translation and rotation around the vertical axis is so similar that it is difficult to distinguish. Because the unique omni-directional camera 360°field of vision can avoid these problems, target detection and tracking based on omni-directional video has been researched in a number of areas, such as video surveillance, robot navigation, and smart driving.This paper analyzes the characteristics of omni-directional image, improves the target detection and tracking algorithm respectively, for the inadequate of classical target detection and tracking algorithm applied to omni-directional images. Research includes the following aspects:(1) Based on the analysis of Markov random field image motion detection principle and the characteristics of omni-directional image resolution and neighborhood, the paper presents a new neighborhood definition method for catadioptric omni-directional image.(2) Based on the analysis of kalman filter image motion tracking theory and the phenomenon that imaging trajectory is conic in the catadioptric omni-directional image when target moves along the straight-line, the paper modifies the kalman filter to improve the state prediction accuracy of target tracking on the catadioptric omni-directional image.Combination of the above methods, the paper verifies the robustness and correctness of algorithms through many experiment videos, and the results are satisfactory.更多还原