VISION ALGORITHMS TO ACQUIRE WORKPIECES FROM A BIN OF PARTS

摘要

Machine feeding is dominated by the use of human labor, mechanical feeders and orientation preservation. These techniques have limitations that make an alternate solution desirable. Vision algorithms that enabled a robot to acquire a single piece from a bin containing randomly placed identical parts were developed. The approach used search for a section of any piece that would enable grasping, with a particular gripper, with a reasonable performance level. The algorithms were hold site driven, and piece and gripper dependent. Two types of grippers were used, vacuum cup and parallel jaw. The hold sites were defined as patches of smooth surfaces for the vacuum cup gripper and opposing parallel edges, linear or curvilinear, for the parallel jaw gripper. The shrinking algorithm used a binary image of the bin. Shrinking the blobs corresponding to the pieces different clusters corresponding to different pieces were identified. Larger blobs were selected as the best hold sites. The collision fronts algorithm extracted edges corresponding to the piece/background separation using a gradient operator. The edge points were propagated in fronts towards the middle of the piece. Opposing propagating edges collided forming collision fronts. Patterns associated with those fronts enabled placement of a parallel jaw gripper. The parallel jaw filter algorithm use the matched filtering technique. Masks taking into account the characteristics of the piece, gripper and noise were applied to the image. Good matches were used to identify the pick up points. To make the use of match filters practical, hold sites were as small as possible, the low-pass component of the filtering was applied early in the process, and an eigen type decomposition was obtained. A single bin image was used. Once a hold site was identified the gripper travelled along the line of sight, the distance from the part to the camera being obtained by proximity sensors in the gripper. An overload sensor, to abort acquisition in case of error, and a grasping sensor, to indicate successful acquisition, were also used. Camera calibration procedures to allow the line of sight travel were developed and are presented in appendix.