基于改进代价计算和自适应引导滤波的立体匹配 
下载： 1405次
Stereo Matching Method Based on Improved Cost Computation and Adaptive Guided Filter
图 & 表
图 1. 算法流程图
Fig. 1. Diagram of proposed method
下载图片 查看原文
图 2. 基于不同代价计算方法的Tsukuba图像的初始视差图。(a)原始梯度代价函数;(b)增强后的梯度代价函数;(c)原始Census变换;(d)基于增强梯度的Census变换
Fig. 2. Initial disparity maps based on different cost methods for Tsukuba. (a) Absolute difference in images gradients; (b) absolute difference in enhanced images gradients; (c) traditional Census transform; (d) Census transformation based on enhanced images gradients
下载图片 查看原文
图 3. 自适应窗口构建示意图。(a)基于交叉的区域结构;(b)文献[
17]的自适应窗口;(c)文献[
14]的自适应窗口;(d)本文方法的自适应窗口
Fig. 3. Schematic of adaptive window construction. (a) Cross-based support region construction; (b) adaptive window in Ref. [17]; (c) adaptive window in Ref. [14]; (d) adaptive window in proposed method
下载图片 查看原文
图 4. Aloe和Baby1在不同光照下不同代价计算方法的视差图。(a)左图像;(b)右图像;(c)真实视差图;(d) SAD和梯度结合;(e) AD和Census变换结合;(f) AD、梯度和Census变换结合;(g)本文算法
Fig. 4. Disparity maps under different illumination conditions for Aloe and Baby1. (a) Left image; (b) right image; (c) ground truth; (d) SAD+Grad; (e) AD+Cen; (f) AD+Grad+Cen; (g) proposed cost computation
下载图片 查看原文
图 5. Aloe和Baby1在不同曝光下不同代价计算方法的视差图。(a)左图像;(b)右图像;(c)真实视差图;(d) SAD和梯度结合;(e) AD和Census变换结合;(f) AD、梯度和Census变换结合;(g)本文算法
Fig. 5. Disparity maps with different exposures for Aloe and Baby1. (a) Left image; (b) right image; (c) ground truth; (d) SAD+Grad; (e) AD+Cen; (f) AD+Grad+Cen; (g) proposed cost computation
下载图片 查看原文
图 6. 不同代价聚合算法在弱纹理图像的视差图。(a)左图像;(b)真实视差图;(c)基于传统引导滤波器的视差图;(d)传统引导滤波器算法的误匹配像素图;(e)本文算法的视差图;(f)本文算法的误匹配像素图
Fig. 6. Disparity maps of different cost aggregation algorithms for textureless images. (a) Left images; (b) ground truth maps; (c) results of local stereo method based on guided filter; (d) error maps for method based on guided filter; (e) results of the proposed method; (f) error maps of the proposed method
下载图片 查看原文
图 7. 不同参数设置的实验结果
Fig. 7. Experimental results on different parameter settings
下载图片 查看原文
表 1实验参数设置
Table1. Experimental parameter settings
| Parameter | Value | | Parameter | Value |
|---|
| λGRAD | | 25 | λCTg | 15 | | τ1 | 30 | | L1 | 31 | | τ2 | 6 | | L2 | 80 | | dLim | 9 | | ε | 0.012 |
|
查看原文
表 2不同代价计算算法在不同光照下的误匹配率
Table2. Error matching rates of various cost computations under different illuminations%
| Algorithm | Aloe | Baby1 | Bowling1 | Cloth1 | Flowerpots | Rocks1 | Avg |
|---|
| SAD+Grad | 32.175 | 16.882 | 40.900 | 10.829 | 53.528 | 27.238 | 30.259 | | AD+Cen | 32.274 | 25.055 | 46.147 | 13.212 | 56.000 | 18.732 | 31.903 | | AD+Grad+Cen | 37.149 | 23.175 | 46.658 | 12.690 | 72.106 | 32.375 | 37.359 | | Proposed | 22.034 | 11.115 | 26.946 | 11.333 | 34.185 | 13.849 | 19.910 |
|
查看原文
表 3不同代价计算算法在不同曝光下的误匹配率
Table3. Error matching rates of various cost computations under different exposures%
| Algorithm | Aloe | Baby1 | Bowling1 | Cloth1 | Flowerpots | Rocks1 | Avg |
|---|
| SAD+Grad | 52.510 | 50.672 | 46.434 | 50.178 | 87.562 | 79.773 | 61.188 | | AD+Cen | 16.173 | 11.118 | 20.022 | 11.096 | 41.021 | 15.329 | 19.127 | | AD+Grad+Cen | 31.012 | 30.182 | 31.374 | 13.543 | 77.590 | 44.218 | 37.987 | | Proposed | 15.205 | 10.658 | 22.782 | 11.060 | 29.834 | 14.094 | 17.272 |
|
查看原文
表 4不同代价计算算法在无幅度失真条件下的误匹配率
Table4. Error matching rates of various cost computations without radiometric changes%
| Algorithm | Aloe | Baby1 | Bowling1 | Cloth1 | Flowerpots | Rocks1 | Avg |
|---|
| SAD+Grad | 12.409 | 12.009 | 26.122 | 9.619 | 20.697 | 10.598 | 15.242 | | AD+Cen | 13.610 | 11.811 | 23.859 | 10.475 | 22.676 | 12.766 | 15.866 | | AD+Grad+Cen | 15.349 | 12.350 | 24.563 | 11.236 | 21.832 | 12.586 | 16.319 | | Proposed | 14.478 | 9.749 | 18.663 | 11.085 | 18.644 | 12.008 | 14.104 |
|
查看原文
表 5不同算法的误匹配率
Table5. Error matching rates of different algorithms for different images%
| Algorithm | Tsukuba | Venus | Teddy | Cones | Avg |
|---|
| n-occ | all | disc | | | n-occ | all | disc | | n-occ | all | disc | | n-occ | all | disc |
|---|
| GF | 2.21 | 2.59 | 8.56 | 0.32 | 0.68 | 4.31 | 4.77 | 8.62 | 13.1 | 2.53 | 7.90 | 7.67 | 5.27 | | Proposed | 1.74 | 1.95 | 8.35 | 0.23 | 0.42 | 3.17 | 3.95 | 7.88 | 10.8 | 2.80 | 8.11 | 8.25 | 4.80 |
|
查看原文
表 6所有区域不同算法的误匹配率
Table6. Error matching rates of different algorithms in all regions%
| Algorithm | Aloe | Baby1 | Baby2 | Baby3 | Bowling1 | Bowling2 |
|---|
| GF | 7.407 | 2.575 | 5.534 | 5.981 | 7.940 | 12.184 | | Proposed | 8.626 | 4.092 | 10.635 | 6.197 | 14.636 | 14.794 | | Algorithm | Cloth1 | Cloth2 | Cloth3 | Cloth4 | Flowerpots | Lampshade1 | | GF | 2.960 | 8.613 | 3.940 | 8.393 | 12.405 | 11.223 | | Proposed | 3.225 | 10.418 | 4.332 | 8.454 | 12.696 | 9.540 | | Algorithm | Lampshade2 | Midd1 | Midd2 | Monopoly | Plastic | Rocks1 | | GF | 15.729 | 37.653 | 35.381 | 22.803 | 32.666 | 4.183 | | Proposed | 8.570 | 13.857 | 16.270 | 7.335 | 25.724 | 4.968 | | Algorithm | Rocks2 | Wood1 | Wood2 | Avg(all) | | | | GF | 3.587 | 3.829 | 0.965 | 11.712 | | | | Proposed | 3.973 | 8.574 | 0.484 | 9.400 | | |
|
查看原文
表 7不同算法在弱纹理图像的误匹配率
Table7. Error matching rates of different algorithms for textureless images%
| Algorithm | Lampshade1 | Lampshade2 | Midd1 | Midd2 | Monopoly | Plastic | Avg |
|---|
| CostFilter | 23.242 | 31.811 | 48.993 | 45.200 | 36.796 | 43.758 | 38.300 | | CS-GF | 10.720 | 8.634 | 29.127 | 25.892 | 14.439 | 22.178 | 18.498 | | CS-MST | 14.955 | 16.360 | 18.294 | 17.496 | 30.626 | 37.933 | 22.610 | | CS-ST | 13.201 | 12.188 | 16.072 | 9.587 | 24.053 | 30.724 | 17.638 | | Proposed | 9.540 | 8.570 | 13.857 | 16.270 | 7.335 | 25.724 | 13.549 |
|
查看原文
表 8不同算法的运行时间比较
Table8. Runtime comparison of different algorithms for benchmark stereo imagess
| Algorithm | Tsukuba | Venus | Teddy | Cones |
|---|
| CostFilter | 1.18 | 2.46 | 6.41 | 6.47 | | CS-GF | 2.76 | 5.12 | 15.07 | 15.55 | | CS-MST | 2.14 | 2.59 | 5.88 | 5.98 | | CS-ST | 1.95 | 2.51 | 5.57 | 5.61 | | Proposed | 3.42 | 5.85 | 14.469 | 14.253 |
|
查看原文
闫利, 王芮, 刘华, 陈长军. 基于改进代价计算和自适应引导滤波的立体匹配[J]. 光学学报, 2018, 38(11): 1115007. Li Yan, Rui Wang, Hua Liu, Changjun Chen. Stereo Matching Method Based on Improved Cost Computation and Adaptive Guided Filter[J]. Acta Optica Sinica, 2018, 38(11): 1115007.
PDF全文
