Automatic Extrinsic Calibration and Workspace Mapping

Abstract

In small to medium enterprises (SME), industrial robot arms are not used very much despite the fact that they offer a large potential to increase the competitiveness. The problem is that to be effective in the SME sector, robot systems should be more flexible. To be economically effective a robot arm should be used with multiple different tasks which are usually in different locations in the company. Therefore, the robot should be able to be reconfigured quickly to the new task and location giving a low robot set-up time. Prevention of collisions with obstacles in the different locations and the extrinsic calibration of sensors on the robot are two important and time consuming tasks during set-up of the robot. Automatic generation of a 3D map of obstacles near the robot would decrease set-up time and therefore increase the flexibility. The calibration of the camera is also a tedious procedure which could be automated. Therefore, in this thesis a system is proposed which extrinsically calibrates a camera on the robot arm, and uses the calibrated camera to scan the environment in a safe way to create a collision map. Several calibration methods are investigated. The method created by Tsai et al. gives the best results and is selected. The calibration system works by estimating the position of the camera relative to a checker-board pattern from multiple points and matching this with the robot orientation. From this the camera position and orientation on the robot are calculated. In this thesis the system is tested in simulation and on a Universal robots UR5 robot arm. To safely map the environment without colliding in to obstacles which are not seen yet a Next Best View (NBV) viewpoint generation system is proposed. The system generates viewpoints for the camera which add the most new information to the map. A virtual wall in the map round the robot represents the unknown space, which prevents the robot to move into unknown space. By viewing the unknown space with the camera the unknown space is cleared. As monocular mapping algorithms are not available yet, a 3D camera is used for data acquisition. Data is stored in an OctoMap system, which is a memory efficient, discretized probabilistic storage system. This system is also implemented on the UR5 robot arm. The calibration system works successfully. In addition, we concluded that noise in the camera pose estimation system is the main limiting factor for achieving high precision. The collision map system has been implemented correctly and also functions successfully. For this part of the system, we conclude that the main limiting factor for the processing speed could be removed if the mapping system would be integrated in the NBV software.