
基于双目视觉的改进特征立体匹配方法
Improved Feature Stereo Matching Method Based on Binocular Vision
针对特征立体匹配方法只能得到稀疏视差以及弱纹理匹配率低以及视差精度不足导致视差连续处不平滑而呈阶梯状等问题,提出一种改进的特征立体匹配算法. 对预处理后的左右图像提取特征并进行特征匹配,再经过筛选获取准确的匹配点对;将得到的稀疏匹配点对作为种子点,依据视差连续性与极线约束准则建立一维搜索空间,利用积分图简化的快速零均值归一化互相关系数作为相似度量函数,通过双向匹配策略实现区域生长,大大提升匹配准确性的同时降低算法复杂度;通过亚像素拟合和加权中值滤波后处理提升视差精度,去除视差阶梯分层、噪声和条纹现象. Middlebury数据集实验结果表明,本算法得到了准确性更高且稠密的视差,提高了弱纹理区与深度不连续处的匹配效果以及整体视差精度,同时具有很强的鲁棒性,能抑制一定亮度差异和噪声的影响.
The feature-based stereo matching method can only get sparse disparity, and the low matching rate of weak texture areas and insufficient disparity accuracy lead to problems such as unsmooth continuity and stepped disparity. An improved features stereo matching algorithm is proposed. Extract features from the left and right images after preprocessing and perform feature matching, and then filter to obtain accurate matching point pairs; The obtained sparse matching point pairs are used as seed points, the search space is established according to the disparity continuity and the extreme line constraint criterion, and the fast zero-mean normalized cross-correlation simplified by integral graph is used as the similarity measurement function to achieve region growth through a two-way matching strategy, which greatly improves matching accuracy and reduces matching complexity; Sub-pixel fitting and weighted median filtering are used to improve the accuracy of disparity, remove disparity step layering, noise and streaks. The experimental results of the Middlebury data set show that this algorithm obtains a highly accurate and dense disparity, improves the matching effect of weak texture areas and depth discontinuities, and the accuracy of disparity. At the same time, it has strong robustness and can suppress the influence of certain brightness differences and noise.
双目视觉 / 立体匹配 / 稠密视差 / 区域生长 / 双向匹配 / 视差精化 {{custom_keyword}} /
binocular vision / stereo matching / dense disparity / regional growth / two-way matching / disparity refinement {{custom_keyword}} /
|
---|
Input:Seed points on the left (xl, yl) and right(xr, yr) Output:Dense disparity d Step: 1 fori = -1 to 1 2 forj = -1 to 1 3 Lx←xl + j 4 Ly←yl + i 5 forjj = -1 to 1 6 Rx←xr + jj 7 Ry←yr+ i 8 Matching cost {(Lx, Ly), (Rx, Ry)} 9 Select the minimum matching cost: mincost 10 if (cost< Threshold &&cost<mincost) then 11 mincost←cost 12 jmin←jj 13 end if |
14 end for 15 Reverse matching (based on the right) 16 Rx1← xr+jmin 17 Ry1← yr+i 18 fork = -1 to 1. //Polar point on the left 19 Lx1← Lx+k 20 Ly1← Ly+i 21 cost1 {(Lx1, Ly1), (Rx1, Ry1)} 22 if (cost1< Threshold &&cost1<mincost1) then 23 mincost1←cost1 24 jmin1←k 25 end if 26 end for 27 Calculate the disparity: d 28 if (costmin < Threshold &&(Lx==xl +jmin1)) then 29 printd←abs(Lx-(xr+jmin)) 30 end for 31 end for |
表1 特征检测与匹配结果 |
Image | Number of FAST feature point before pretreatment | Number of FAST feature point after pretreatment | Number of matching point pairs before/after Eliminate false matches | |||
---|---|---|---|---|---|---|
Left | Right | Left | Right | |||
Venus | 536 | 526 | 627 | 637 | 78 | 74 |
Bull | 221 | 223 | 384 | 340 | 40 | 40 |
Cones | 515 | 519 | 640 | 587 | 54 | 53 |
Teddy | 349 | 380 | 383 | 403 | 43 | 41 |
表2 其他算法与本文算法实验图像结果对比 |
Image | Objective index | NCC | BM | SGBM | DP | Proposed algorithm |
---|---|---|---|---|---|---|
Venus | Error rate | 0.179 | 0.219 | 0.056 | 0.275 | 0.020 |
Bull | Error rate | 0.099 | 0.166 | 0.050 | 0.179 | 0.014 |
Cones | Error rate | 0.236 | 0.324 | 0.193 | 0.368 | 0.131 |
Teddy | Error rate | 0.286 | 0.369 | 0.232 | 0.452 | 0.148 |
表3 改变右图亮度的实验结果 |
Image | Objective index | 20 cd/m2 | 30 cd/m2 | 40 cd/m2 | 50 cd/m2 | Original image |
---|---|---|---|---|---|---|
Sawtooth | Correct rate | 96.43% | 96.43% | 96.35% | 96.21% | 97.32% |
1 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
2 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
3 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
4 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
5 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
6 |
王阳萍, 秦安娜, 郝旗, 等. 结合加速鲁棒特征的遥感影像半全局立体匹配[J]. 光学学报, 2020, 40(16): 163-171.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
7 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
8 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
9 |
肖进胜, 田红, 邹文涛, 等. 基于斜平面平滑优化的半全局立体匹配[J]. 电子学报, 2018, 46(8): 1835-1841.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
10 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
11 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
12 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
13 |
MEN Y,
{{custom_citation.content}}
{{custom_citation.annotation}}
|
14 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
15 |
毛昕蓉, 王楠. 基于ADCensus的改进双目立体匹配算法[J].电子产品世界, 2020, 27(4): 62-64, 87.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
16 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
17 |
欧阳鑫玉, 张娟娟, 赵楠楠, 等. 基于NCC的改进立体匹配算法[J]. 微型机与应用, 2015, 34(3): 54-57.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
18 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
19 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
20 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_ref.label}} |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
/
〈 |
|
〉 |