Robotik 1 - Kapitel 9 - Bildverarbeitung in der Robotik (Teil 3)

In this talk

1. 1:28

   Canny edge detection

   Canny aims to find edges accurately as thin, well located pixel lines. Its stages include Gaussian noise reduction, gradient calculation, non maximum suppression, thresholding, and tracking connected edges with hysteresis. In a kitchen scene, the method detects strong contours on objects such as a cereal box, juice carton, and cups, though lighting can affect which edges appear.

2. 19:26

   Visual servoing for manipulation

   Visual servoing uses camera data to correct robot motion as it happens, accounting for imperfect kinematic models and uncertain object localization. Position based methods compare the hand’s estimated 3D pose with the object’s pose, while image based methods drive visual features toward desired locations in the image. The lecture connects camera motion to feature motion through an image Jacobian and describes its use in grasping with the AMA robot.

3. 54:50

   Point cloud normals and registration

   For a local point cloud neighborhood, PCA estimates a surface normal from the eigenvector associated with the covariance matrix’s smallest eigenvalue. Registration aligns views in a shared coordinate frame; ICP repeatedly pairs each point with its nearest neighbor in another cloud and estimates a rotation and translation. ICP works best when the views overlap and can converge to a local minimum.

4. 1:08:50

   RANSAC and geometric models

   RANSAC estimates shapes such as lines in 2D or planes in 3D by repeatedly sampling the minimum points needed for a model and counting nearby inliers. It is simple and tolerates outliers, but its results are nondeterministic and can be unreliable when inliers are scarce relative to outliers.

5. 1:15:48

   Scene affordances for humanoid robots

   The research goal is to interpret scenes by the actions their objects and surfaces allow, such as leaning on vertical surfaces, supporting weight on horizontal ones, or lifting and pushing objects. The pipeline combines RGB-D perception, point cloud segmentation and registration, geometric primitives, grasp hypotheses, and reachability maps. Robots then test suggested interactions, since geometric evidence alone cannot confirm that an object will actually move or support weight.