midterm presentation

[surberj]

Main Content:

• Motivation:
  • Inspection / maintenance task,
  • GPS-based localization not accurate enough
  • VI Odometry not accurate enough (drift)
• Proposed Approach with localization against Reference Map
  • VI Odometry + Correction based on matches between current Frames and Reference Map
  • Setup for master thesis: flight1 --> offline creation of reference map --> flight2
  • flow diagram of proposed approach
• VI-Odometry results on Unispital and Ticino datasets
  • Issue of Drift
  • General applicability of VI-odometry in the large
  • Uncertainty way higher than conventional EKF with IMU, airspeed, GPS, baro, magnetometer
  • Issue with small baseline
• Reference Map (offline)
  • Representation as point cloud with associated BRISK descriptors
  • Representation with keyframes with BRISK descriptors
  • Representation with keyframes with BRISK descriptors and associated landmarks
  • Closing the loop (Place recognition) to resolve drift
  • Run global BA to resolve drift
• Register new frame with respect to Reference Map
  • 2D2D matching with BRISK features
  • Unispital datasets: Robustness under viewpoint changes
  • 2D3D matching with BRISK features
  • I will not be able to evaluate this until the midterm
• Estimation of base frame transformation based on structure matches
• Outline Roadmap

[surberj]

Todo's: 2D2D matching with BRISK features on Unispital dataset:

• [ ] 2D2D matching with opencv::matcher()
  • [ ] Implementation (should be more or less straight forward)
  • [ ] Analysis: number of inlier-matches over "distance" between frames
• [ ] 2D2D matching with okvis::DenseMatcher()
  • [ ] Implementation (templates version, should be faster than cv::matcher())
  • create a new class Matcher, which is subclass of MatchingAlgorithm and does the matcher-setup and calls the DenseMatcher
  • create a new class MatchToReference which handles the data flow
  • create a new class ReferenceMap which stores at least all keyframes as multiframes and their pose T_WS (follow Marlins KeyframeDatabase class)
  • [ ] Analysis: number of inlier-matches over "distance" between frames

Reference Map:

• [ ] Create Visualizations of OKVIS map (Unispital + Ticino)
• [ ] Comparison to offline tools? --> VisualSFM, Theia-Ceres
• [ ] Comparison to ROVIO?

Presentation:

• [ ] create ppt