opencv - 不准确的实时视差图。

Question

我是立体相机的新手(或者可能是 opencv 的新手)。但是，我需要做我的 FYP，需要立体相机来测量所选对象的距离。

获得准确视差图的步骤是什么？据我所知，我们需要校准以获得内在和外在参数，对其进行去畸变，然后对其进行校正，然后只计算视差映射。

我已经完成校准以获得内在和外在值(value)。然后我试图纠正它，我得到了这个 RMS value_and_reprojection_error_value.jpg .问题是，视差映射的可接受值是多少？

之后，我开始使用 Block Matching 进行视差映射。这是我的代码`

int main(void)
{

   VideoCapture camLeft(0);
   VideoCapture camRight(2);

   camLeft.set(CV_CAP_PROP_FRAME_WIDTH, 500);
   camLeft.set(CV_CAP_PROP_FRAME_HEIGHT, 500);
   camRight.set(CV_CAP_PROP_FRAME_WIDTH, 500);
   camRight.set(CV_CAP_PROP_FRAME_HEIGHT, 500);



   if (!camLeft.isOpened() || !camRight.isOpened()) {
      cout << "Error: Stereo Cameras not found or there is some problem       connecting them. Please check your cameras.\n";
      exit(-1);
     }


//Read intrinsice parameters
string intrinsic_filepath = "C:/Users/Jerry/Documents/Visual Studio 2015/Projects/OctStereoCalibration/OctStereoCalibration/intrinsics.yml";
FileStorage fs(intrinsic_filepath, FileStorage::READ);
if (!fs.isOpened())
{
    printf("Failed to open intrinsics.yml");
    return -1;
}
Mat M1, D1, M2, D2;
fs["M1"] >> M1;
fs["D1"] >> D1;
fs["M2"] >> M2;
fs["D2"] >> D2;

//Read Extrinsic Parameters
string extrinsic_filepath = "C:/Users/Jerry/Documents/Visual Studio 2015/Projects/OctStereoCalibration/OctStereoCalibration/extrinsics.yml";
fs.open(extrinsic_filepath, FileStorage::READ);
if (!fs.isOpened())
{
    printf("Failed to open extrinsics");
    return -1;
}

Mat R, T, R1, P1, R2, P2;
fs["R"] >> R;
fs["T"] >> T;

Mat frame1, frame2, gray1, gray2, copyImageLeft, copyImageRight;
int counter = 0;

camLeft >> frame1;
camRight >> frame2;

Size img_size = frame1.size();
Rect roi1, roi2;
Mat Q;

stereoRectify(M1, D1, M2, D2, img_size, R, T, R1, R2, P1, P2, Q, CALIB_ZERO_DISPARITY, -1, img_size, &roi1, &roi2);

Mat map11, map12, map21, map22;
initUndistortRectifyMap(M1, D1, R1, P1, img_size, CV_16SC2, map11, map12);
initUndistortRectifyMap(M2, D2, R2, P2, img_size, CV_16SC2, map21, map22);



while (1) {
    createTrackbars();
    on_trackbar(0, 0);

    bm->setROI1(roi1);
    bm->setROI2(roi2);
    bm->setPreFilterCap(PreFilterCap);
    bm->setPreFilterSize(PrefilterSize);
    bm->setBlockSize(SADWindowSize);
    bm->setMinDisparity(MinDisparity);  //0
    bm->setNumDisparities(numberOfDisparities);
    bm->setTextureThreshold(TextureThreshold);
    bm->setUniquenessRatio(UniquenessRatio);
    bm->setSpeckleWindowSize(SpeckleWindowSize);
    bm->setSpeckleRange(SpeckleRange);
    bm->setDisp12MaxDiff(Disp12MaxDiff);    //1

    camLeft >> frame1;
    camRight >> frame2;
    if ((frame1.rows != frame2.rows) || (frame1.cols != frame2.cols)) {
        cout << "Error: Images from both cameras are not of some size. Please check the size of each camera.\n";
        exit(-1);
    }
    //frame1.copyTo(copyImageLeft);
    //frame2.copyTo(copyImageRight);
    imshow("Cam1", frame1);
    imshow("Cam2", frame2);

    /************************* STEREO ***********************/

    cvtColor(frame1, gray1, CV_RGB2GRAY);
    cvtColor(frame2, gray2, CV_RGB2GRAY);

    int64 t = getTickCount();

    Mat img1r, img2r;
    remap(gray1, img1r, map11, map12, INTER_LINEAR);
    remap(gray2, img2r, map21, map22, INTER_LINEAR);

    Mat disp, disp8;
    Mat XYZ;
    bm->compute(img1r, img2r, disp);
    t = getTickCount() - t;
    printf("Time elapsed: %fms\n", t * 1000 / getTickFrequency());

    disp.convertTo(disp8, CV_8U, 255 / (numberOfDisparities*16.));
    //normalize(disp, disp8, 0, 255, CV_MINMAX, CV_8U);
    imshow("disparity", disp8);

    //reprojectImageTo3D(disp8, XYZ, Q, false, CV_32F);

    char keyBoardInput = (char)waitKey(50);
    if (keyBoardInput == 'q' || keyBoardInput == 'Q') {
        break;
        return(0);
    }
}

}`

我的 BM 参数是:


int PreFilterCap = 31;
int PrefilterSize = 21;
int SADWindowSize = 33;
int 最小差异 = 0;
int numberOfDisparities = 48;
int 纹理阈值 = 29;
int 唯一性比率 = 15;
int SpeckleWindowSize = 32;
int SpeckleRange = 32;
int Disp12MaxDiff = 0;
我得到的视差图是: Disparity map

如何获得质量更好的视差映射？谢谢

Answer 1

您可以尝试更改参数窗口大小和视差数量的值。对于不同的深度，不同的值集有效。
第二种选择是使用 SGBM 而不是 sbm，我知道的第三种方法是使用高斯金字塔来计算左右相机在低分辨率下的视差，然后将它们与一些简单的逻辑融合。在低层级计算视差图的好处是，当您沿着金字塔层级往下走时，孔将开始填充，从而使最终的视差图更好。