Using Active Vision for Enhancing an Surface-based Object Recognition Approach

For robots, the sensing of their surroundings is a necessary skill in nowadays tasks. One possible realization are one-shot object recognition methods. These may fail due to occlusions within the scene or because of objects that cannot be identified from only one view. This problem may be tackled by utilizing Active Vision methods, which means capturing additional information of the scene from different poses. In this paper, we contribute a novel approach with the goal of increasing the performance of an object recognition method. To do this, we try to identify so called regions of interest that the robot should inspect. We describe, how regions of interest can be determined and how possible views are calculated based on the current representation of the scene. On the one hand, we use the resulting additional views to generate new object hypotheses. On the other hand, they enable the validation of existing hypotheses. We test for each possible view whether it is reasonable regarding distance to the scene and occlusions. For remaining views, the corresponding pose as well as a quality measure is calculated and the best pose is selected as the next view. Evaluation results from our prototype system show that the classification rate increases with additional views.