6DOF Object Recognition and Positioning for Robotics Using Next Best View Heuristics

Abstract

The accuracy and portability of depth cameras have increased by a lot in recent years, which allows for advanced 3D scanning of the environment for robotic applications. In this thesis we have developed a system that uses a depth camera mounted on a robot arm to identify and localize arbitrary objects, and give hints on how to move the camera to get better localization results. The system works by generating virtual views of objects to identify them in a point cloud generated by the depth camera. This data is then used to estimate a pose of the object, and generate a hint on where to move the camera next. After a new point cloud is taken, it is merged with the previous cloud which allows the system to iteratively get more confident in the identification and pose of the object.