A Semi-autonomous Approach to Object Pose Estimation based on Object Templates

Avatar Robots play an important role in dangerous environments such as post-disaster scenarios that are not accessible for humans. In order to make these robots more prac- ticable in the future, a fast interaction between the robot and a human supervisor that solves high-level planning tasks is required. One of these planning tasks is the recogni- tion of objects in the sensor data of the environment to enable the robot to interact with the object. As it is a time consuming part of this task to provide the pose (consisting of translation, rotation and scale) of a given object template to the robot, a semi-autonomous approach to estimate the pose of a user-specified object template is presented in this thesis. To reduce the interaction of the human supervisor to a minimum, the provided algorithm operates only with a specified template of an arbitrary object and a rough initial position of the object in the environment of the robot. The provided algorithm also has to deal with different challenges such as partial overlap of the template with the object in the sensor data, noise and occlusions. It is also required to be easily extendable, for example for fast tracking of an object if the pose of the object in the past is known. Furthermore an interface is provided that keeps the effort of the human supervisor to provide the necessary information to the object alignment algorithm to a minimum and allows for quick interaction and feedback