Information driven sensor placement for robust active object recognition based on multiple views

Abstract

Robust object recognition is a mandatory prerequisite for many applications in Service Robotics. A common approach is to capture a single image of the scene from a fixed position and to recognize all objects at once. This challenging task is even more demanding in everyday environments. We propose an approach for object recognition which makes use of mobile manipulation. An initial object belief is enhanced by systematically inspecting objects from different views with a second camera on a manipulator. Knowledge about an object is used to generate possible viewpoints that are directed at specific features. Viewpoint selection considers the expected recognition probability and costs to optimize the recognition performance. The approach is demonstrated in real-world experiments with a service robot. Almost identically appearing objects are reliably classified by systematic inspection of additional distinguishing features like barcodes and text labels.