Deformation-Based Loop Closure for Large Scale Dense RGB-D Slam

The inventors present a system for capturing large scale dense maps in an online setting with a low cost RGB-D sensor. This has significant applications in robotics, where a robot’s ability to create a map of an unknown environment and know its position within that map is crucial for intelligent autonomous operation.  

Researchers

Michael Kaess / John McDonald / Thomas Whelan / John Leonard

Departments: Department of Mechanical Engineering
Technology Areas: Industrial Engineering & Automation: Autonomous Systems, Robotics / Sensing & Imaging: Optical Sensing
Impact Areas: Connected World

  • a method for mapping an environment¿
    United States of America | Granted | 9,412,173

Technology      

The Inventors’ SLAM system is comprised of two main components, the frontend (for camera tracking and surface extraction) and the backend (for pose graph and map optimization). The frontend used is the Kintinuous mapping system, based on the KinectFusion™,  system. This real-time dense mapping system integrates all depth measurements into a volumetric data structure and carries out camera pose estimation is then carried out via a dense Iterative Closest Point (ICP) function. As the volumetric structure moves, the region of the surface that leaves the volume is extracted in point cloud form. The stream of “cloud slices” along a camera trajectory each have an associated camera pose. The final component in the frontend is a visual place recognition module that detects visual loop closures and computes appropriate relative camera pose constraints.The backend performs incremental pose graph optimization coupled with incremental non-rigid dense map optimization. A dense every-frame pose graph is optimized according to the loop closure constraints. The optimized dense camera trajectory is used to conjunction with matched visual features to constrain a non-rigid space deformation of the map. The inventors use an embedded deformation approach to apply it to large scale dense maps captured with the Kintinuous frontend and derive efficient incremental methods to prepare the map for deformation. By combining pose graph optimization with non-rigid deformation of a dense map, they obtain highly accurate dense maps over large scale trajectories that are both locally and globally consistent.  

Problem Addressed 

Simultaneous Localization and Mapping (SLAM) in 3D space is a key research area in modern robotics. Notable RGB-D 3D reconstruction systems include KinectFusion™, which enables dense volumetric modeling of a static scene in real-time at sub-centimeter resolution. Current solutions to adapt this technology over extended areas suffer from problems of unbounded drift due to an open-loop process. The Inventors present a method to address loop closures in a volume shifting-based mapping systems that take advantage of camera pose graph optimization and non-rigid space deformation. The result is a visual SLAM system that captures high fidelity dense maps in real-time with the local reconstruction quality of KinectFusion™,  and the advantages of local consistency given by camera pose graph optimization.  

Advantages 

  • Low-cost and easy-to-implement system based on existing RGB-D technology and frameworks.
  • Incremental method for deformation graph constructions allows multi-million point maps to be captured over hundreds of meters in real-time.
  • Highly accurate dense maps over large scale trajectories are both locally and globally consistent.  

Publications

"Deformation-Based Loop Closure for Large Scale Dense RGB-D SLAM." In Intelligent Robots and Systems (IROS), 2013, November 3–7, 2013.

"Real-Time Large-Scale Dense RGB-D SLAM with Volumetric Fusion." International Journal of Robotics Research, December 9, 2014, pp. 598–626.

License this technology

Interested in this technology? Connect with our experienced licensing team to initiate the process.

Sign up for technology updates

Sign up now to receive the latest updates on cutting-edge technologies and innovations.