双目立体视觉技术及其硬件实现研究

发布时间：2018-04-23 18:06

本文选题：局部立体匹配 + 图像分割　；参考：《南京理工大学》2017年硕士论文

【摘要】：双目立体视觉是计算机视觉的重要研究方向,利用双目立体视觉可以获取三维场景的深度信息,已经广泛应用在军用与民用的各个领域。双目立体视觉的关键在于立体匹配,因此成为目前该领域的主要研究内容。本文重点研究了如何提高局部立体匹配算法在深度不连续区域以及低纹理区域的匹配精度,同时搭建了基于FPGA的立体匹配以及显著目标测距系统。研究了基于图像分割的局部立体匹配算法,针对算法在深度不连续区域的匹配精度不高以及对亮度差异敏感等缺陷,提出并研究了一种基于非参数变换和图像分割的高效聚合立体匹配算法。该算法以分割块作为支持窗口,以Census变换后的序列值作为匹配基元,采用本文提出的匹配代价函数进行代价计算。同时提出了一种动态视差范围矫正的技术,以邻域像素视差值为参考对当前像素的视差搜索范围进行调整,提高匹配效率。实验结果证明该算法能有效解决深度不连续区域的误匹配问题。研究了基于低纹理区域的局部立体匹配算法,针对低纹理区域缺少特征差异性,提出并研究了一种基于像素颜色空间和窗口位置的立体匹配算法。该算法首先对图像中的低纹理区域进行检测,然后采用本文提出的基于像素颜色空间的匹配代价函数进行代价计算,最后根据像素在聚合窗口中的不同位置分配相应的权值并将代价值聚合起来。实验结果证明该算法可以有效提高低纹理区域的匹配精度。搭建了一套以FPGA为图像处理核心的双目立体视觉系统。利用FPGA内部的并行计算和流水线设计,探索并设计了一种基于Box滤波的方法来实现Census变换;探索并设计实现了通过图像锐化和阈值分割来提取红外图像中的显著目标以及测距;设计实现了对左右视频流进行立体显示。实验结果表明,该硬件系统能够实时对图像进行立体匹配和显著目标提取,近距离的测距精度较高,系统的稳定性较好。
[Abstract]:Binocular stereo vision is an important research direction of computer vision. Using binocular stereo vision to obtain depth information of 3D scene has been widely used in military and civil fields. Stereo matching is the key of binocular stereo vision, so it has become the main research content in this field. This paper focuses on how to improve the matching accuracy of local stereo matching algorithm in deep discontinuous region and low texture region. At the same time, a stereo matching system based on FPGA is built. The local stereo matching algorithm based on image segmentation is studied. An efficient aggregate stereo matching algorithm based on nonparametric transformation and image segmentation is proposed and studied. The algorithm takes the partition block as the supporting window and the sequence value of Census transform as the matching primitive. The proposed matching cost function is used to calculate the cost. At the same time, a technique of dynamic disparity range correction is proposed, in which the parallax range of the current pixel is adjusted to improve the matching efficiency by using the parallax value of neighboring pixels as a reference. Experimental results show that the algorithm can effectively solve the problem of mismatch in depth discontinuous region. The local stereo matching algorithm based on low texture region is studied, and a stereo matching algorithm based on pixel color space and window position is proposed and studied for the lack of feature difference in low texture region. The algorithm firstly detects the low-texture region in the image, and then calculates the cost by using the matching cost function based on the pixel color space proposed in this paper. Finally, the corresponding weights are assigned according to the different positions of pixels in the aggregation window and the generation values are aggregated. Experimental results show that the algorithm can effectively improve the matching accuracy of low texture regions. A binocular stereo vision system with FPGA as the core of image processing is set up. Using the parallel computing and pipeline design in FPGA, this paper explores and designs a method based on Box filtering to realize the Census transform, and explores and designs to extract the prominent targets and ranging from infrared images by image sharpening and threshold segmentation. The stereoscopic display of left and right video streams is designed and realized. The experimental results show that the hardware system can carry out stereo matching and salient target extraction in real time, and the precision of ranging in close range is higher, and the stability of the system is better.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】