快速光流估计与时空一致件三维重建

发布时间：2018-04-12 21:04

本文选题：光流估计 + 大位移　；参考：《浙江大学》2017年硕士论文

【摘要】：光流是计算机视觉中的重要研究领域,在运动分割、物体识别、目标跟踪、视频差值、三维重建等方面均有应用。光流估计问题是计算机视觉中一个经典和基础的问题。自从光流被提出之后,大量研究人员和学者开始从事光流的计算方面的研究。近年来专家学者们陆陆续续提出了很多有效的光流估计方法。目前光流估计仍然存在一些问题,如在实际工程应用中面临大位移和户外场景时难以得到较为满意的结果,计算复杂度非常高,时耗大,即使使用GPU实现也很难做到实时估计。本文提出了一种针对大位移问题的一种非常有效的快速光流估计方法,该方法能够有效并快速的得到较为准确的光流估计结果。该方法将匹配思想与传统变分优化方法相结合,使之能够良好地处理大位移情况下的光流估计问题。结合特征匹配去噪,增加可靠的匹配点数量,进而快速地获得稠密的初始化光流场。在初始化光流场的基础上,使用基于变分优化的细节优化,实现了高质量快速的稠密光流场估计。此外,我们还采用了 GPU实现对算法进行了加速,从而能满足一些对光流计算速度要求较高的应用要求。同时将我们的光流估计算法应用于三维重建,在现有方法上增加了时空一致性优化改善了重建结果。利用连续帧之间的关联信息,建立了连续帧间的光滑约束进行时空一致性深度优化,提高了深度恢复和三维重建质量。实验表明,本文的方法在KITTI数据集上有良好的表现。能够在确保一定精度的前提下,进下快速的光流估计。在原有深度结果的基础上,使用我们的光流结果对其进行时空一致性优化,得到了更为准确的深度信息和三维重建结果。
[Abstract]:Optical flow is an important research field in computer vision, which has been applied in motion segmentation, object recognition, target tracking, video difference, 3D reconstruction and so on.Optical flow estimation is a classical and fundamental problem in computer vision.Since the optical flow was proposed, a large number of researchers and scholars began to study the calculation of optical flow.In recent years, many effective optical flow estimation methods have been proposed by experts and scholars.At present, there are still some problems in optical flow estimation, for example, it is difficult to obtain satisfactory results when facing large displacement and outdoor scenes in practical engineering applications, the computational complexity is very high, the time consumption is very large, and it is difficult to achieve real-time estimation even using GPU.In this paper, a very effective fast optical flow estimation method for large displacement problem is proposed. This method can get more accurate results of optical flow estimation effectively and quickly.This method combines the idea of matching with the traditional variational optimization method, so that it can deal with the problem of optical flow estimation in the case of large displacement.Combined with feature matching denoising, the number of reliable matching points is increased, and the dense initial optical flow field is obtained quickly.In addition, we use GPU to speed up the algorithm, which can meet the requirements of some applications with high speed of optical flow calculation.At the same time, our optical flow estimation algorithm is applied to 3D reconstruction, and the spatio-temporal consistency optimization is added to the existing methods to improve the reconstruction results.Using the correlation information between successive frames, the smooth constraints between successive frames are established to optimize the spatio-temporal consistency depth, which improves the quality of depth recovery and 3D reconstruction.Experiments show that the proposed method has a good performance on KITTI data sets.Under the premise of ensuring certain precision, fast optical flow estimation can be achieved.On the basis of the original depth results, we use our optical flow results to optimize the temporal and spatial consistency, and obtain more accurate depth information and 3D reconstruction results.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】