的OpenCV>>相机相对姿态估计

时间:2014-04-16 15:39:42

标签: c++ algorithm opencv computer-vision

环境

  1. 使用3DS Max生成的3D场景
  2. 相机视野为45度
  3. 使用相同的相机@ 800x600分辨率
  4. 渲染两张图像
  5. 图像A具有相机Z旋转== 0度
  6. 图像B具有相机Z旋转== 25度
  7. 8个对应点(手动驱动),无异常值

  8. 手头的任务

    解决图像A和图像A之间的相对相机姿势。期望图像B(上图)在Z轴上产生诱导25度旋转


    实施

    选项A:

    1. 手动生成通信,保证没有异常值(在下面的代码片段中查看'rotZ0'&'rotZ25 ')
    2. 相机焦距以像素分辨率为基础,使用图像分辨率& FOV符合this link
    3. 相机内在矩阵由(this link组成)由图像分辨率和相机FOV组成
    4. 基本矩阵是使用'cv :: findFundamentalMat'
    5. 派生的
    6. 基本矩阵由相机内在矩阵'K'和基本矩阵'F'组成(根据this link)以下方式:'Kt()* F * K'*其中'Kt()'是内在矩阵转置。
    7. 对基本矩阵'matE'
    8. 执行SVD
    9. 解决4种可能的解决方案:[U * W * Vt],[U * W.t()* Vt],[U * W.t()* Vt.t()]& [U * W * Vt.t()]
    10. 选项B:

      1. 手动生成通信,保证没有异常值(参见'rotZ0'&'rotZ25 ')
      2. Essential Matrix使用'cv :: findEssentialMat'
      3. 组成
      4. 使用'cv :: recoverPose'
      5. 估算相机姿势

        结果

        以上两个选项都无法正确恢复相对相机姿势(预计在Z轴上旋转25度)

        我做错了什么?
        如何正确分辨相机的相对姿势?

        任何帮助都将不胜感激。


        完整代码

        #define RAD2DEG(rad) (((rad) * 180)/M_PI)
        #define DEG2RAD(deg) (((deg) * M_PI)/180)
        #define FOV2FOCAL(pixelssensorsize, fov) ((pixelssensorsize) / (2 * tan((fov) / 2)))// http://books.google.co.il/books?id=bXzAlkODwa8C&pg=PA48&lpg=PA48&dq=expressing+focal+length+in+pixels&source=bl&ots=gY4972kxAC&sig=U1BUeNHhOHmYIrDrO0YDb1DrNng&hl=en&sa=X&ei=45dLU9u9DIyv7QbN2oGIDA&ved=0CGsQ6AEwCA#v=onepage&q=expressing%20focal%20length%20in%20pixels&f=false
        
        // http://nghiaho.com/?page_id=846
        void DecomposeRotation(IN const cv::Mat& R, OUT float& fX, OUT float& fY, OUT float& fZ) {// Taken from MatLab
            fX = (float)atan2(R.at<double>(2, 1), R.at<double>(2, 2));
            fY = (float)atan2(-R.at<double>(2, 0), sqrt(R.at<double>(2, 1)*R.at<double>(2, 1) + R.at<double>(2, 2)*R.at<double>(2, 2)));
            fZ = (float)atan2(R.at<double>(1, 0), R.at<double>(0, 0));
        }
        
        int _tmain(int argc, _TCHAR* argv[])
        {
            // 25 deg rotation in the Z axis (800x600)
            const cv::Point2f rotZ0[] = { { 109, 250 }, { 175, 266 }, { 204, 279 }, { 221, 253 }, { 324, 281 }, { 312, 319 }, { 328, 352 }, { 322, 365 } };
            const cv::Point2f rotZ25[] = { { 510, 234 }, { 569, 622 }, { 593, 278 }, { 616, 257 }, { 716, 303 }, { 698, 340 }, { 707, 377 }, { 697, 390 } };
            const cv::Point2f rotZminus15[] = { { 37, 260 }, { 106, 275 }, { 135, 286 }, { 152, 260 }, { 258, 284 }, { 248, 324 }, { 266, 356 }, { 260, 370 } };
        
        
            const double        dFOV = DEG2RAD(45);
            const cv::Point2d   res(800, 600);
            const cv::Point2d   pntPriciplePoint(res.x / 2, res.y / 2);
            const cv::Point2d   pntFocal(FOV2FOCAL(res.x, dFOV), FOV2FOCAL(res.y, dFOV));
        
            //transfer the vector of points to the appropriate opencv matrix structures
            const int                numPoints = sizeof(rotZ0) / sizeof(rotZ0[0]);
            std::vector<cv::Point2f> vecPnt1(numPoints);
            std::vector<cv::Point2f> vecPnt2(numPoints);
        
            for (int i = 0; i < numPoints; i++) {
                vecPnt2[i] = rotZ0[i];
                //vecPnt2[i] = rotZminus15[i];
                vecPnt1[i] = rotZ25[i];
            }
        
            //// Normalize points
            //for (int i = 0; i < numPoints; i++) {
            //  vecPnt1[i].x = (vecPnt1[i].x - pntPriciplePoint.x) / pntFocal.x;
            //  vecPnt1[i].y = (vecPnt1[i].y - pntPriciplePoint.y) / pntFocal.y;
        
            //  vecPnt2[i].x = (vecPnt2[i].x - pntPriciplePoint.x) / pntFocal.x;
            //  vecPnt2[i].y = (vecPnt2[i].y - pntPriciplePoint.y) / pntFocal.y;
            //}
        
            try {
                // http://docs.opencv.org/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html
                cv::Mat matK = cv::Mat::zeros(3, 3, CV_64F);
                matK.at<double>(0, 0) = pntFocal.x;
                matK.at<double>(1, 1) = pntFocal.y;
                matK.at<double>(0, 2) = pntPriciplePoint.x;
                matK.at<double>(1, 2) = pntPriciplePoint.y;
                matK.at<double>(2, 2) = 1;
        
                float x, y, z;
                cv::Mat R1, R2, R3, R4;
                cv::Mat t;
                cv::Mat matE;
        
        #if 1 // Option [A]
                cv::Mat matF = cv::findFundamentalMat(vecPnt1, vecPnt2);
                matE = matK.t() * matF * matK; // http://en.wikipedia.org/wiki/Essential_matrix
        
                cv::Mat _tmp;
                cv::Mat U;
                cv::Mat Vt;
        
                cv::SVD::compute(matE, _tmp, U, Vt);
        
                cv::Matx33d W(0, -1, 0,
                              1,  0, 0,
                              0,  0, 1);
        
                R1 = U*cv::Mat(W)*Vt; // See http://stackoverflow.com/questions/14150152/extract-translation-and-rotation-from-fundamental-matrix for details
                R2 = U*cv::Mat(W)*Vt.t();
                R3 = U*cv::Mat(W).t()*Vt;
                R4 = U*cv::Mat(W).t()*Vt.t();
        #else // Option [B] 
                matE = cv::findEssentialMat(vecPnt1, vecPnt2, pntFocal.x, pntPriciplePoint);// http://docs.opencv.org/trunk/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html
                cv::decomposeEssentialMat(matE, R1, R2, t);
                int iInliers = cv::recoverPose(matE, vecPnt1, vecPnt2, R4, t);// , pntFocal.x, pntPriciplePoint);
                R3 = cv::Mat::zeros(3, 3, CV_64F);
        #endif
        
                DecomposeRotation(R1, x, y, z);
                std::cout << "Euler Angles R1 (X,Y,Z): " << RAD2DEG(x) << ", " << RAD2DEG(y) << ", " << RAD2DEG(z) << std::endl;
                DecomposeRotation(R2, x, y, z);
                std::cout << "             R2 (X,Y,Z): " << RAD2DEG(x) << ", " << RAD2DEG(y) << ", " << RAD2DEG(z) << std::endl;
                DecomposeRotation(R3, x, y, z);
                std::cout << "             R3 (X,Y,Z): " << RAD2DEG(x) << ", " << RAD2DEG(y) << ", " << RAD2DEG(z) << std::endl;
                DecomposeRotation(R4, x, y, z);
                std::cout << "             R4 (X,Y,Z): " << RAD2DEG(x) << ", " << RAD2DEG(y) << ", " << RAD2DEG(z) << std::endl;
        
                //cv::Mat res = matFrom.t() * matF * matTo;// Results in a null vector ( as it should ) http://en.wikipedia.org/wiki/Fundamental_matrix_(computer_vision)
                //res = matFrom.t() * matE * matTo;// Results in a null vector ( as it should )
            }
            catch (cv::Exception e) {
                _ASSERT(FALSE);
            }
            return 0;
        }
        

        执行结果

        选项A:

        Euler Angles R1 (X,Y,Z): -26.2625, 8.70029, 163.643
                     R2 (X,Y,Z): 16.6929, -29.9901, -3.81642
                     R3 (X,Y,Z): 5.59033, -20.841, -19.9316
                     R4 (X,Y,Z): -5.76906, 7.25413, -179.086
        

        选项B:

        Euler Angles R1 (X,Y,Z): -13.8355, 3.0098, 171.451
                     R2 (X,Y,Z): 2.22802, -22.3479, -11.332
                     R3 (X,Y,Z): 0, -0, 0
                     R4 (X,Y,Z): 2.22802, -22.3479, -11.332
        

2 个答案:

答案 0 :(得分:2)

首先,校准相机而不是使用预定义的值。它总是会产生很大的影响。 由8-pt或5-pt计算的相对姿势会受到很多噪音的影响,绝不意味着最终的结论。话虽如此,重建点然后捆绑调整整个场景是个好主意。优化你的外在性,你应该想出一个更好的姿势。

答案 1 :(得分:0)

我认为你的结果都不等于90度是因为坐标系。将摄像机绕z轴旋转90度的坐标系由您自己定义。您可以将相机姿势想象为指向场景的矢量,并且从基本矩阵分解的R矩阵表示矢量的旋转。结果是在另一个坐标系中。您可以通过检查两个摄像机矢量是否具有90度的矢量角来检查我的假设。