Skip to content

Latest commit

 

History

History
31 lines (25 loc) · 1.33 KB

these_rosinol.md

File metadata and controls

31 lines (25 loc) · 1.33 KB

Densifying Sparse VIO: a Mesh-based approach using Structural Regularities

Author: Toni Rosinol

Year: 2018

Front End:

  • Classical Harris detector, KLT tracker, RANSAC, Keyframe selector
  • 3D mesh generator : 2D Mesh to 3D mesh with filtering faces that do not represent an actual surface
  • Regularities detector: landmark to plane data association
  • 2D triangulation prevent from doing a 3D tetrahedralisation that is computationnally expensive
  • 3 pb :
    • noisy positions of lmks
    • irregular density of the point cloud
    • point cloud is morphing, points are removed and added at each iteration
  • Only keypoints that have a stereo corespondence and a 3D landmark are kept for mesh
  • 3D Mesh update:
    • at each frame a 2D Mesh is created from scratch
    • check for every duplicate triangle wrt the last 2D and delete them from the current mesh
    • convert into 3D mesh and append it to the current 3D mesh
    • remove any rows with marginalized landmarks
  • Mesh filtering:
    • pb of triangles from walls to ground and outliers
    • remove silver triangles (triangles with two very accute angles)
    • remove triangles with a large ratio between largest and smallest side
    • remove triangles with an edge larger than a thresh (?)
  • Uses the mesh normals and elevation histogram to extract horizontal planes (table and ground)