针对树型立体视觉匹配的视差预测层级加速算法研究

发布时间：2018-01-06 07:29

本文关键词：针对树型立体视觉匹配的视差预测层级加速算法研究　出处：《上海交通大学》2015年硕士论文　论文类型：学位论文

【摘要】：立体视觉是目前计算机视觉领域一个重要的课题。它的主要目的在于提取场景中的深度信息,利用立体匹配算法得到精准的视差图,进而完成场景的三维重构。基于树型结构的匹配是其中重要的立体匹配算法之一。近年来提出的树型立体视觉匹配算法通过在图像中构造最小生成树(MST)来聚合匹配代价。该类算法结合了高效率局部匹配与高精度全局匹配的特点,匹配精度以及匹配效率较传统匹配算法均有很大提高。其时间复杂度理论值为O(N·dmax),即运行时间只和像素点总数N和图像最大视差dmax相关。但树型匹配算法基于贪心搜索策略,对于构造匹配窗口缺乏一定的自适应性,在同一平面存在不同深度分布的情况下无法得到精确的深度信息。同时,该类算法运行效率受到最大视差限制,在最大视差较大的实际场景和高分辨率图像中,运行速度较慢。本论文针对传统树型立体视觉匹配算法中出现的问题,引入视差预测模型(Disparity Prediction Model,DPM),提出针对树型立体视觉匹配的视差预测层级加速算法(Disparity-Prediction-Based Hierarchical Accelerator,DPA)。该算法通过预测视差概率分布,构造出更精确的核心视差置信区间,使得匹配代价的计算,代价聚合,视差计算,视差优化都只发生在这一核心视差置信区间内,从而降低树型匹配算法的时间复杂度。本论文通过大量的实验数据,包括Middlebury数据库的实验室图像,KITTI数据库的实际场景的高分辨率图像,以及自主拍摄的低质量的实际场景图像,验证了DPA加速算法在提升匹配精度的同时,能够优化基于树型结构匹配算法的复杂度,大大提高运行效率。本论文共有三个主要贡献:1.提出了核心视差置信区间概念。对于每一个像素点,其视差值处于核心视差置信区间范围内的概率远远大于置信区间范围外。通过精确的视差置信区间,可以在匹配代价计算、视差计算以及视差修复等步骤中大大减少计算量,提高运行效率。2.提出了基于视差分布的DPM视差预测模型。此模型在图像金字塔结构基础上,通过小规模视差预测出大规模视差的概率分布,并应用于构造核心视差置信区间。3.提出了普适的针对树型立体视觉匹配的DPA层级加速算法。此算法可结合到现有的树型匹配算法框架中。经实验验证,该算法大大提高了原始树型匹配算法的运行效率。图像分辨率越高,视差范围越大,效果越明显。具体地,对于高分辨率图像,应用了DPA层级加速算法后,匹配效率提升达7-10倍。
[Abstract]:Stereo vision is an important subject in the field of computer vision. The main purpose of stereo vision is to extract the depth information from the scene and use stereo matching algorithm to get accurate parallax images. The matching based on tree structure is one of the important stereo matching algorithms. In recent years, the tree stereo vision matching algorithm is proposed by constructing the minimum spanning tree MSTs in the image. This algorithm combines the characteristics of high efficiency local matching and high precision global matching. The accuracy and efficiency of matching are greatly improved compared with the traditional matching algorithm, and the theoretical value of time complexity is OGN 路dmax. That is, the running time is only related to the total number of pixels N and the maximum parallax dmax of the image. However, the tree matching algorithm is based on greedy search strategy, so it has no self-adaptability for constructing matching window. In the case of different depth distribution in the same plane, the accurate depth information can not be obtained. At the same time, the efficiency of this algorithm is limited by the maximum parallax, in the actual scene with large maximum parallax and in the high-resolution image. In this paper, the disparity Prediction Model is introduced to solve the problems in the traditional tree-type stereo vision matching algorithm. DPM. A hierarchical acceleration algorithm for disparity prediction based on tree-type stereo vision matching is proposed. Disparity-Prediction-Based Hierarchical Accelerator. By predicting the parallax probability distribution, the algorithm constructs a more accurate confidence interval of the core parallax, which makes the calculation of matching cost, cost aggregation and parallax calculation. Parallax optimization only occurs in this core disparity confidence interval, thus reducing the time complexity of the tree matching algorithm. It includes laboratory images of Middlebury database and high-resolution images of actual scenes in KITTI database, as well as low-quality actual scene images taken independently. It is verified that the DPA acceleration algorithm can improve the matching accuracy and optimize the complexity of the tree-based matching algorithm. In this paper, there are three main contributions: 1. The concept of the core parallax confidence interval is proposed. For each pixel point, the core parallax confidence interval is proposed. The probability that the apparent difference is in the confidence range of the core parallax is much larger than that in the confidence interval. Through the accurate parallax confidence interval, the matching cost can be calculated. In the steps of parallax calculation and parallax repair, the computation amount is greatly reduced, and the running efficiency is improved. 2. A parallax distribution based DPM disparity prediction model is proposed. The model is based on the image pyramid structure. The probability distribution of large parallax is predicted by small parallax. And it is applied to construct the confidence interval of core parallax. 3. A universal DPA hierarchical acceleration algorithm for tree stereo vision matching is proposed. This algorithm can be combined into the existing framework of tree matching algorithm. The algorithm greatly improves the efficiency of the original tree matching algorithm. The higher the image resolution, the larger the parallax range, the more obvious the effect. Specifically, for high-resolution images, the DPA hierarchical acceleration algorithm is applied. Matching efficiency increases by 7-10 times.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】