Skip to content

Latest commit

 

History

History
19 lines (15 loc) · 933 Bytes

a_new_visual_front_end.md

File metadata and controls

19 lines (15 loc) · 933 Bytes

A New Visual Front-end Combining KLT with Descriptor Matching for Visual-inertial Odometry

Author: Wu

Year: 2023

Notes:

  • Tightly coupled KLT and descriptor matching for indirect VIO. Test method on both filter based and optimization based algorithm. Use KLT and superpoint (CNN method for extraction)
  • Tackles the pb of lost tracked that are rediscovered after
  • complementary strength: tracker for small motion, descriptor for big
  • run tracking and matching in parallel and perform validation
  • SuperPoint: Following the encoder, the architecture is split into two decoders, each designed to learn the weights of specific tasks: one for keypoint detection and the other for descriptor computation
  • KLT module: forward / backward tracking
  • Descriptor module: KNN algorithm with k = 2 + forward backward tracking
  • Points that are succesfully tracked with both
  • Then perform RANSAC filtering with fundamental matrix