自然场景图像的光流场估计

Estimation of optical flow field of nature scene images

自然场景图像序列中,物体运动会造成部分背景区域的显露和遮挡,显露和遮挡区域的像素在连续相邻的图像中缺乏对应点,因而传统的光流方法常常在这些区域给出错误的光流估计。图像在采集、传输过程中可能会受到噪声污染,噪声干扰是进行光流场估计必须考虑的另外一个重要问题。为消除显露、遮挡和噪声干扰引起的光流估计误差,采用新的可视矩阵标记图像位置的遮挡、显露、可视三种状态,以此来引导光流场估计,并采用正态概率分布对图像噪声的分布状态进行近似,从而在Bayes框架下建立了自然场景图像光流场估计的数学模型,最后通过迭代方法获得了致密的光流场。采用CAVIAR视频数据对本文算法进行测试并与Negal光流法进行性能对比,结果表明,本文方法具有更好的光流场估计效果。

In nature scene image sequences, a covered-uncovered problem occurs when objects cover one part of background areas and uncover another because of motion. Pixels of covered-uncovered areas will lose counterpoints in consecutive images, thus the traditional optical flow methods often determine a bad optical flow at these pixels. Noise disturbance is another key problem that should be taken into account in optical flow estimation, because image data may be blurred in capture and transmission process. To avoid the determining errors arising from covered-uncovered and noise disturbance problems, a visible matrix is adopted to label the three possible states of image areas, namely cov- ered/uncovered/visible, and a normal probability distribution is adopted to approximately describe the distribution of noises. Based on these, the arithmetic expressions of the optical flow field are deduced in the Bayes framework, and a dense optical flow field is determined using the iteration method, then CAVIAR video data are used to test the proposed algorithm, and a comparison is made between the experiment results and Negal＇s method. The results show that the proposed algorithm performs better.