(Mur-Artal 2017) VI-ORB

Search IconIcon to open search

URL: http://ieeexplore.ieee.org/abstract/document/7817784
Authors: Mur-Artal, Tardós
Code: http://paperswithcode.com/paper/visual-inertial-monocular-slam-with-map-reuse
Results (video): http://www.youtube.com/watch?v=JXRCSovuxbA

Abstract

  • current VI odometry approaches: drift accumulates due to lack of loop closure
  • therefore there is a need for tightly-coupled VI-SLAM with loop closure and map reuse
  • here: focus on monocular case, but applicable to other camera configurations
  • builds on ORB-SLAM (from same author)
  • IMU initialisation method (initialises: scale, gravity direction, velocities, gyroscope bias, accelerometer bias) depends on visual monocular initialisation (coupled initialisation)

Other works: recent tightly-coupled VIO (both filtering- and optimisation-based) lack loop closure, so drift accumulates

Introduction

Why use the visual-inertial sensor combination?

Framework of VI-ORB

3 parallel threads

  1. [Front end] Tracking Tracking via optimisation of the current frame, assuming a fixed map (unchanged map)

  2. [Back end] Local mapping

    • local BA over several keyframes (sliding window)
    • this local BA is a compromise between
      • full smoothing (over all keyframes) — high computational complexity
      • marginalising out past keyframes (loss of information)

  3. Loop closure

Initialisation

  • Need for a reliable VI initialisation that provides accurate state estimate
  • Because both the tracking (front end) and BA (back end) fix the states in their optimisations, this can bias the solution –> need for initialisation  fix states?: states which aren’t the argument of the optimisation function, i.e. fixed (not optimised)
  • Optimal solution for initialisation of all the required init. variables [scale, gravity, biases, velocities, structure of the opt. graph, camera pose] would require a full BA,
    • however this is split into smaller steps
  • The proposed initialisation is general and applicable to any keyframe-based monocular SLAM
  • Requirement: any two consecutive keyframes must be close in time (to reduce IMU noise integration)
  1. Process the first few seconds of video with visual monocular SLAM (here using ORBSLAM)
    • This gets the structure estimate as well as several keyframe poses scaled by an unknown scale
    • Use a motion that makes all variables observable
  2. Compute gravity bias from orientation of keyframes
  3. Initial guess for scale, accelerometer bias (using known magnitude of gravity 9.81 m/s2)
  4. Refine the scale and gravity direction
  5. Get velocities for all keyframes

When reinitialising after relocalisation (after a long time; using place recognition):

  • Reinitialise bg gyrometer bias
  • Scale s and gravity g already known from first initialisation, so no need to calculate anew
  • Estimate ba accelerometer bias from the same equation used during initialisation (simplified now due to knowledge of s, g)