Exploitation of 3D Information for Directing Visual Attention and Object Recognition

Abstract

In the field of vision based robot actuation, in order to find and grasp objects in its environment, object recognition is a fundamental task that should be carried out in a fast and efficient way. Although a high resolution imaging environment is convenient for object recognition, it has high computational and memory costs. In this paper, we present a novel method to segment possible object locations using a low resolution range camera so that the further object recognition step, by using a high resolution camera, is only performed in a few candidate regions. A coarse 3D representation of the whole environment is obtained by stitching the range images while the arm is exploring the scene. Exploiting the obtained 3D information of the environment and the objects, a saliency map is built. The high resolution camera mounted on the robot arm is directed to the candidate (salient) regions in the saliency map for more detailed analysis. Finally, object recognition is performed using scale invariant features in the high resolution images. 1