转自https://blog.csdn.net/cxgincsu/article/details/74451940
立体匹配算法对比研究
引言:基于OpenCV3.0,对BM、SGBM和GC算法进行了对比测试研究。由于SGBM算法视差效果好速度快的特点,常常被广泛应用和改进,本文针对SGBM算法主要参数设置作了对比测试,以供大家参考。
BM:Block Matching ,采用SAD方法计算匹配代价;
SGBM: 修改自Heiko Hirschmuller的《Stereo Processing by Semi-global Matching and Mutual Information》;
与原方法不同点:
- 没有实现原文中基于互信息的匹配代价计算,而是采用BT算法("Depth Discontinuities by Pixel-to-Pixel Stereo" by S. Birchfield and C. Tomasi);
- 默认运行单通道DP算法,只用了5个方向,而fullDP使能时则使用8个方向(可能需要占用大量内存);
- 增加了一些BM算法中的预处理和后处理程序;
GC:OpenCV3.0中没有实现,可以在OpenCV以下版本中找到。该方法效果是最好的,但是速度太慢,不能达到实时的匹配效率;
1、SGBM
主要参数:minDisparity 、numDisparities、blockSize、P1、P2。其他参数设置参照http://blog.csdn.net/zhubaohua_bupt/article/details/51866567
代码:
-
#include "stdafx.h"
-
#include "opencv2/opencv.hpp
-
using namespace std;
-
using namespace cv;
-
-
int _tmain(int argc, _TCHAR* argv[])
-
{
-
Mat left = imread("imgL.jpg
", IMREAD_GRAYSCALE);
-
Mat right = imread("imgR.jpg
", IMREAD_GRAYSCALE);
-
Mat disp;
-
-
int mindisparity = 0;
-
int ndisparities = 64;
-
int SADWindowSize = 11;
-
-
//SGBM
-
cv::Ptr<cv::StereoSGBM> sgbm = cv::StereoSGBM::create(mindisparity, ndisparities, SADWindowSize);
-
int P1 = 8 * left.channels() * SADWindowSize* SADWindowSize;
-
int P2 = 32 * left.channels() * SADWindowSize* SADWindowSize;
-
sgbm->setP1(P1);
-
sgbm->setP2(P2);
-
-
sgbm->setPreFilterCap(15);
-
sgbm->setUniquenessRatio(10);
-
sgbm->setSpeckleRange(2);
-
sgbm->setSpeckleWindowSize(100);
-
sgbm->setDisp12MaxDiff(1);
-
//sgbm->setMode(cv::StereoSGBM::MODE_HH);
-
-
sgbm->compute(left, right, disp);
-
-
disp.convertTo(disp, CV_32F, 1.0 / 16); //除以16得到真实视差值
-
Mat disp8U = Mat(disp.rows, disp.cols, CV_8UC1); //显示
-
normalize(disp, disp8U, 0, 255, NORM_MINMAX, CV_8UC1);
-
-
imwrite("results/SGBM.jpg
", disp8U);
-
return 0;
-
}
minDisparity:最小视差,默认为0。此参数决定左图中的像素点在右图匹配搜索的起点,int 类型;
numDisparities:视差搜索范围长度,其值必须为16的整数倍。最大视差 maxDisparity = minDisparity + numDisparities -1;
blockSize:SAD代价计算窗口大小,默认为5。窗口大小为奇数,一般在3*3 到21*21之间;
P1、P2:能量函数参数,P1是相邻像素点视差增/减 1 时的惩罚系数;P2是相邻像素点视差变化值大于1时的惩罚系数。P2必须大于P1。需要指出,在动态规划时,P1和P2都是常数。
一般建议:P1 = 8*cn*sgbm.SADWindowSize*sgbm.SADWindowSize;P2 = 32*cn*sgbm.SADWindowSize*sgbm.SADWindowSize;
tsukuba_left.jpg | tsukuba_right.jpg |
SADWindowSize = 5 | SADWindowSize = 11 | SADWindowSize = 21 |
P1 = 8 , P2 = 10 | P1 = 8 , P2 = 32 | P1 = 8 , P2 = 64 |
MODE_SGBM(5方向) | MODE_HH(8方向) |
总结:
2. 惩罚系数控制视差图的平滑度,P2>P1,P2越大则视差图越平滑;
3. 八方向动态规划较五方向改善效果不明显,主要在图像边缘能够找到正确的匹配;
2、BM
代码:
-
#include "stdafx.h"
-
#include "opencv2/opencv.hpp"
-
using
namespace
std;
-
using
namespace cv;
-
-
int _tmain(
int argc, _TCHAR* argv[])
-
{
-
Mat left = imread(
"imgL.jpg", IMREAD_GRAYSCALE);
-
Mat right = imread(
"imgR.jpg", IMREAD_GRAYSCALE);
-
Mat disp;
-
-
int mindisparity =
0;
-
int ndisparities =
64;
-
int SADWindowSize =
11;
-
-
cv::Ptr<cv::StereoBM> bm = cv::StereoBM::create(ndisparities, SADWindowSize);
-
// setter
-
bm->setBlockSize(SADWindowSize);
-
bm->setMinDisparity(mindisparity);
-
bm->setNumDisparities(ndisparities);
-
bm->setPreFilterSize(
15);
-
bm->setPreFilterCap(
31);
-
bm->setTextureThreshold(
10);
-
bm->setUniquenessRatio(
10);
-
bm->setDisp12MaxDiff(
1);
-
-
copyMakeBorder(left, left,
0,
0,
80,
0, IPL_BORDER_REPLICATE);
//防止黑边
-
copyMakeBorder(right, right,
0,
0,
80,
0, IPL_BORDER_REPLICATE);
-
bm->compute(left, right, disp);
-
-
disp.convertTo(disp, CV_32F,
1.0 /
16);
//除以16得到真实视差值
-
disp = disp.colRange(
80, disp.cols);
-
Mat disp8U = Mat(disp.rows, disp.cols, CV_8UC1);
-
normalize(disp, disp8U,
0,
255, NORM_MINMAX, CV_8UC1);
-
imwrite(
"results/BM.jpg", disp8U);
-
return
0;
-
}
SADWindowSize = 5 | SADWindowSize = 21 | SADWindowSize = 31 | SADWindowSize = 41 |
3、GC
OpenCV3.0暂未实现GC算法,这里直接贴出结果(来自:http://www.cnblogs.com/polly333/p/5130375.html)
总结:
SGBM | BM | GC |
视差效果:BM < SGBM < GC;
处理速度:BM > SGBM > GC ;
参考:
http://www.cnblogs.com/polly333/p/5130375.html
http://www.cnblogs.com/polly333/p/5130375.html
http://blog.csdn.net/u014493244