Learning discriminative representations from RGB-D video data
IJCAI'13 Proceedings of the Twenty-Third international joint conference on Artificial Intelligence
| Hi-index | 0.00 |
The advent of RGB-D cameras which provide synchronized range and video data creates new opportunities for exploiting both sensing modalities for various robotic applications. This paper exploits the strengths of vision and range measurements and develops a novel robust algorithm for localization using RGB-D cameras. We show how correspondences established by matching visual SIFT features can effectively initialize the generalized ICP algorithm as well as demonstrate situations where such initialization is not viable. We propose an adaptive architecture which computes the pose estimate from the most reliable measurements in a given environment and present thorough evaluation of the resulting algorithm against a dataset of RGB-D benchmarks, demonstrating superior or comparable performance in the absence of the global optimization stage. Lastly we demonstrate the proposed algorithm on a challenging indoor dataset and demonstrate improvements where pose estimation from either pure range sensing or vision techniques perform poorly.