The result of our pose estimation algorithm tested on 12 different configurations of 3 car parts. The measurements are shown as point clouds, and the detected components are shown in red, green and blue.

In many production lines, the parts that are produced arrive with random position and orientation on a conveyor belt. To be able to pick the parts off the conveyor belt and place them with known orientation in bins, it is necessary to recognize the position and pose of all of the objects. This information can in turn be used by the robot to systematically pick and place the objects.

We use a structured light camera to capture 3D images of the scene and we apply 3D image analysis algorithms to determine the position and pose of the objects on the conveyor belt with 6 degrees of freedom. As input to the algorithm for pose determination, we use a template in the form of a point cloud. This template can be obtained by performing a 3D scan of the object, or it might be obtained by sampling points on the surface of a CAD model of the object.

We initialize the detection by performing range image segmentation using depth discontinuities in the scene combined with a measure of range data quality. We initialize the pose detection by detecting geometric primitives in the object and thus removing many degrees of freedom. We further refine the object pose estimate by exhaustibly searching through the remaining degrees of freedom using a scaled-down model, and by measuring distance between segmented range data and the model. This provides us with starting points for final fitting by ICP.

We believe that the described method is applicable for a wide range of industrial automation problems where precise localization of 3D objects in a scene is needed.

Related publications

• J. Thielemann, Ø. Skotheim, J. O. Nygaard and T. Vollset. "System for conveyor belt part picking using structured light and 3D pose estimation". Proc. SPIE Vol. 7239 (2009)

For more information, please contact Øystein Skotheim.