Package boofcv.alg.geo.pose

Class PoseFromPairLinear6

java.lang.Object
boofcv.alg.geo.pose.PoseFromPairLinear6
public class PoseFromPairLinear6 extends Object

Estimates the camera motion using linear algebra given a set of N associated point observations and the depth (z-coordinate) of each object, where N ≥ 6. Note this is similar to, but not exactly the PnP problem.

Output from this class is a rotation and translation that converts a point from the first to second camera's reference frame:

X' = R*X+T
where R is a rotation matrix, T is a translation matrix, X is a coordinate in 1st reference frame, and X' in the second.

This approach is a modified version of the approach discussed in [1]. It is derived by using bilinear and trilinear constraints, as is discussed in Section 8.3. It has been modified to remove redundant rows and so that the computed rotation matrix is row major. The solution is derived from the equations below and by computing the null space from the resulting matrix:
cross(x2)*(A*x1) + cross(x2)*T/λi=0
where cross(x) is the cross product matrix of X, xi is the pixel coordinate (normalized or not) in the ith image, A is rotation and T translation.

[1] Page 279 in "An Invitation to 3-D Vision, From Images to Geometric Models" 1st Ed. 2004. Springer.

  • Constructor Details

    • PoseFromPairLinear6

      public PoseFromPairLinear6()

  • Method Details

    • process

      public boolean process(List<AssociatedPair> observations, List<Point3D_F64> locations)
      Computes the transformation between two camera frames using a linear equation. Both the observed feature locations in each camera image and the depth (z-coordinate) of each feature must be known. Feature locations are in calibrated image coordinates.
      Parameters:
      observations - List of observations on the image plane in calibrated coordinates.
      locations - List of object locations. One for each observation pair.

    • processHomogenous

      public boolean processHomogenous(List<AssociatedPair> observations, List<Point4D_F64> locations)
      Computes the transformation between two camera frames using a linear equation. Both the observed feature locations in each camera image and the depth (z-coordinate) of each feature must be known. Feature locations are in calibrated image coordinates.
      Parameters:
      observations - List of pixel or normalized image coordinate observations
      locations - List of object locations in homogenous coordinates. One for each observation pair.

    • getProjective

      public DMatrixRMaj getProjective()
      P=[A|T]
      Returns:
      projective A

    • getA

      protected DMatrixRMaj getA()
      Matrix used internally.