Reinforcement Learning of Object Detection Strategies

Abstract. Mobile agents performing dynamic sensing without control on information acquisition rely on arbitrarily distributed information. The integration of irrelevant, ambiguous or misleading evidence may result in poor classification performance. In contrast, active fusion schemes seek to acquire evidence that is most appropriate to the current task, e.g., to disambiguate the current state of belief. This paper understands the fusion process in visual object detection as a sequential decision problem. Reinforcement learning enables to develop an efficient strategy to fuse decisive information in terms of a sensorimotor mapping. The presented system learns object models from visual appearance and uses a connectionist architecture for a probabilistic interpretation of the 2-D views. The expected gain in the global classification accuracy provides a utility measure to reinforce actions leading to discriminative viewpoints. The system is verified in experiments with a sewer robot on the task of visually detecting house inlets in sewage pipes for navigation purposes. Crucial improvements in performance are gained using the learned fusion strategy in contrast to arbitrary action selections.