Triangulation

Recovering a point’s 3D coordinates from two cameras’ known positions and where that point lands in each image.

Triangulation is a basic geometric operation in computer vision: a 3D point projects to one pixel in each of two or more images, and given each camera’s projection matrix (determined by its intrinsics and extrinsics), a ray is drawn from each camera’s optical center through the corresponding image point; the intersection of these rays is the point’s 3D location. In practice, lens distortion and feature-localization error mean the rays usually don’t intersect exactly, so the best estimate is instead found with the midpoint method, a linear solve (DLT), or by minimizing reprojection error — the gap between the observed point and the 3D point projected back into the image. Stereo depth sensing, structure from motion (SfM), and visual SLAM mapping all rely on triangulation to turn 2D matched points into 3D points.

ExampleA stereo camera matches the same corner of a cup’s rim in its left and right images; given both lenses’ intrinsics and relative pose, triangulation computes how far that corner is from the camera.

Related

Epipolar Geometry · Camera Intrinsics · Camera Extrinsics · Stereo Camera · Structure from Motion · Reprojection Error

Sources

Triangulation (computer vision) - Wikipedia