Towards Active Object-Based Navigation

Abstract

This thesis investigates the design and implementation of an active object-based navigation system. Localization refers to the capability of a mobile robot to determine its position relative to a prior map of the environment. Localization based on fiducial markers, such as AprilTags, has been a popular technique used widely in many laboratories for over a decade; however, it is desirable to enable a mobile robot to navigate based on the objects that occur naturally in a given environment. The goal of using real objects as “geometric beacons” for robot navigation was identified many years ago. In practice, however, it has been challenging to develop metrically accurate robot localization systems that use semantic, object-based maps. Recent years have seen tremendous progress in the use of machine learning techniques to recognize objects from camera images, including 6 degree-of-freedom (DOF) object pose estimation. Using these machine learning capabilities for robot navigation is an important application, yet research in creating real-time object-based localization systems has been limited. One of the difficulties is that to obtain accurate 6DOF object pose estimates, typically the object needs to be centered in the field of view of the camera, and viewed from a favorable viewing distance. To help meet this need, this thesis has developed a pan-tilt actively controlled camera mount, and deployed it on a small mobile robot, to explore capabilities for improved object-based navigation. By running a detector in real-time as the robot navigates through the world, the active pan-tilt head can actively track an object of interest; 6DOF pose estimates from the object are used to estimate the robot’s pose. The performance of the system is analyzed using ground truth provided by an OptiTrack motion capture system. Future work will extend the system to track multiple objects and to perform active SLAM.