Robust Object Detection for Service Robotics

Abstract

In the near future robotics will become a much larger part of our society, and more robots will be developed to execute a range of gen- eral tasks, so that they are able to assist human operators in their day to day work. At the Technical University of Eindhoven (TU/e), the AMIGO robot is designed to perform such general service tasks, with a special focus on health care applications. A general purpose service robot, such as AMIGO, needs a certain skill set to perform tasks autonomously. One of these skills, is being able to perceive ob- jects in the world. Using this information AMIGO can act accordingly, e. g., by bringing the required object back to its operator. This thesis describes a system for detecting household objects in domestic environments. The object detection system has been devel- oped for AMIGO, and is based on the object detection method, LINE- MOD, presented by hinterstoisser et al. [1]. To make the system appli- cable for both the general service tasks performed by AMIGO and the RoboCup competition, an annual event encouraging robotic research, LINE-MOD is extended by adding extra color modalities to the detec- tion system, testing different normal extraction routines and enabling the use of additional features. The detection module is subsequently included in a framework with a different set of tools, making it of practical use for the AMIGO project and its users, and is integrated with existing systems used currently used in the AMIGO project. The detection module is evaluated with household objects that AMIGO encounters while performing service related tasks and competing in the RoboCup competition. The system is discovered to be able to recognize most household objects with reasonable accuracy, the use of an added color modality results in the increased accuracy of the detection of household objects at a small performance penalty. However, it may still fail in situations where objects are highly textured, resulting in false positives inside object instances, and when object models are too similar in shape or color a problem which is primarily caused by the use of highly quantized feature types.