Loading...

Visual Odometry using RGB-D Cameras

Mohammadi Kaji, Mahsa | 2014

513 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 45820 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Kasaei, Shohreh
  7. Abstract:
  8. Vision-based localization and 3D orientation estimation of a moving camera, has been for long a vast research area including robot localization and mapping, virtual reality and structure from motion. By introduction of RGB-D cameras in 2010, many sparse methods which are based on key-point extraction and tracking, moved towards dense methods. Dense methods utilize the RGB-D depth and gray-scale values in the images and define the odometry estimation problem as an image registration optimization, without the need to make key-point correspondance in images. Although RGB-D cameras impose specific constraints such as limited depth, depth errors and medium resolution, dense methods have shown nontheless successful in odometry estimation of indoor sequences.In this work, an efficient and accurate dense visual odometry framework for indoor odometry estimation of RGB-D cameras is proposed. The proposed method is developed over the keyframe based Inverse Compostional registration (IC) to achieve higher accuracy and efficiency. We propose the theoretical concepts of image alignments and prove that as long as the keyframe is not updated the precomputed Jacobians remain constant in the IC registration method. Therefore the IC alignment is adopted to achieve both accuracy and efficiency. We utilize keyframes to reduce odometry drift. To update key-frames, the View Overlap measure is proposed. We also benefit from a pixel weighting function to robustify the method against occlusion, slight dynamic change and image noise.The results of the proposed method have been evaluated on the publicly known RGB-D database of the Technical University of Munich. Results show an average of 28% improvement in RMSE of translational drift compared to the methods that use other registrations methods and do not benefit from key-frames. Runtime statistics also show an average of 31% improvement in precomputing Jacobian
  9. Keywords:
  10. Visual Odometry ; RGB-D Camera ; Inverse Compositional ; Jacobian Precomputation

 Digital Object List

 Bookmark

...see more