Introduction

CoconutSLAM is a SLAM written in bad C++ code during a holiday. It is bad and the code is terrible, but it is a SLAM.

General

• Shi-Tomasi corner detection
• Orb compute descriptors
• BFMatcher match points
• extract R, t from Essential matrix
• triangulate based on R, t
• a thread for pangolin

Results

Library

• opencv
  • image processing, feature extraction, matching features
• pangolin
  • 3d visualisation
• g2o coming soon
  • graph optimisation, bundle adjustment

Todo

• refactor code in standard directories (include, src, bin...)
• use other data structure to store points (dictionary or map?) to avoid vector + complex indexing
• g2o optimization
• save and load point maps
• filter points behind camera
• use better matching algorithms (done, HAMMING 2)
• if it doesn't work, we use ORB and implement modifications in ORB SLAM 2
• essential matrix is weird, why
• triangulate isn't great
• is there a way to estimate intrinsic matrix without additional information
• more test set (drone, driving, indoor)

Todo (Far future)

• generate mesh using points
• deep learning models to regenerate 3d structure based on points
• deep learning models to filter?