Evaluation of Kinect vision sensor for bin-picking applications: Improved component separation accuracy with combined use of depth map and color image

This report describes a problem involved with use of Kinect depth maps for robot picking of randomly stacked components, and also a solution to this problems. When Kinect is installed above stacked parts and processing is performed using only the obtained Kinect depth map information, there are cases when individual small metal components cannot be separately identified. So that the robot can reliably pick up a single component in these cases, this report demonstrates that in areas where the system incorrectly identifies multiple components as a single component, the addition of color image information and blob analysis of the color image results in accurate separation of the individual components, allowing a single item to be identified for picking.