Place recognition with an attentive robot

Abstract

This thesis is concerned with robotic place recognition. From the robot’s perspective, the goal of place recognition is to determine its whereabouts - given its knowledge of places it has visited previously. This problem is broken into four stages: First, the problem of how the robot should look around in the current place is considered and a model of attentive visual behaviours based on a family of artificial potential functions is developed so that the robot is able to consider the interesting points in the scene. Next, the problem of place representation is considered. For this, a topological model referred to as bubble space is developed. Sensory features associated with all the interesting points are encoded as bubble surfaces - in a manner that is implicitly dependent on robot pose. For compact representation, bubble surfaces are transformed into bubble descriptors that are rotationally invariant with respect to heading changes while being computable in an incremental manner as each new set of visual observations is made. Bubble descriptors are used for supervised place learning and recognition with support vector machines. Following, the problem of integration of sensory and semantic cues is considered and a representation referred to as “cue descriptor semantic forest” is proposed. The integration of different feature and semantic cues becomes very simple regardless of their type or their number while allowing relative and possibly partial semantic expressiveness via cue weights. Finally, the problem of unsupervised place learning is considered. A novel approach for the incremental and autonomous construction of bubble descriptor semantic trees is proposed. Comparative experiments on benchmark datasets as well as real-time robotic experiments indicate high recognition rates, dynamic with low computation times.