外极面图像的运动遮挡模型和运动纹理方向检测算法

KINETIC OCCLUSION MODELING AND MOTION ORIENTATION DETECTION IN EPIPOLAR PLANE IMAGES

运动遮挡边界处的运动估计是一个困难的问题．外极面图像方法将运动估计转化为轨迹线的检测．人造物体的轨迹线容易通过边缘跟踪的方法获得，但对于纹理复杂的自然景物，轨迹跟踪较为困难．外极面图像中的时空纹理（或称之为运动纹理）从长时序列来看整体方向性非常强，这种方向性表达了物体的深度信息．这就意味着可通过大尺度的时空纹理的方向检测来估计深度，通过该纹理的宽窄来分割不同的物体．本文在分析时空频域运动遮挡模型的基础上，提出的运动方向估计算法，结合大窗口频域方向检测和时空域方向确认和精化，较好地兼顾运动方向精度和运动边界定位的问题，同时避免了特征对应和轨迹跟踪．该算法为三维全景建模的核心算法．

Motion estimation at kinetic occlusion boundary is a hard problem in motion analysis. The epipolar plane image method simplifies the problem as locus detection in 2D spatio temporal (ST) images if the motion is a pure translation, but the edge tracking approach proved to be problematic for natural scene images. However, the ST textures (or motion textures), though not identical along the motion direction for a natural scene, present strong orientations when they are observed globally across a long image sequence. The oriented motion textures represents the depth information of the scene. This paper presents a kinetic occlusion model under piecewise linear motion assumption in ST domain and frequency domain. A depth estimation algorithm is given which combines a large Gaussian windowed Fourier transform method with the spatio temporal orientation identification and refinement. In this way both the accuracy of motion orientation and the occlusion boundary localization are guaranteed. Feature extraction, feature matching and locus tracking are avoided. This algorithm is an important part for the 3D panoramic scene modeling.