Learning-Deep-Learning

CubeSLAM: Monocular 3D Object SLAM

June 2020

tl;dr: Monocular SLAM with 3D MOD.

Overall impression

This is an extension to orb-slam by tracking higher level objects rather than key points alone. The name seems to be a play on QuadricSLAM RAL 2017 where each object is represented by an ellipsoid.

It also has a way to perform single frame 3D object detection using 2D bbox and computer vision. Recent advances in DL based mono 3DOD can benefit this aspect.

The SLAM part integrated the size and location of the cuboid into the factor graph optimization. Also it uses motion model to constrain possible movement of cuboids. This way the velocity is also part of the factor graph optimization. This is to be compared with the more flexible DL approach in struct2depth.

Most existing SLAM methods assume the environment to be static or mostly static. Dynamic SLAM is the way to go for autonomous driving.

Key ideas

Technical details

Notes