Simultaneous Localisation and Mapping (SLAM) has been a long-standing research topic in robotics. However, SLAM used to focus on representing geometry (e.g. maps consisting of point clouds), but did not care about individual objects.
In 2018, initial work on QuadricSLAM was published. In QuadricSLAM, objects are the most important parts of a map. We represent each object as a constrainted dual quadric, essentially an ellipsoid in 3D space.
Using this representation is interesting, because it can be represented with only 9 parameters (3 for the shape, 3 for translation, and 3 for rotation). More importantly, it showed how to constrain these parameters from the bounding boxes produced by a deep-learned object detector.
In follow-up work, these ideas were expanded adding initialisation based on depth data, support plane constraints, and integration with a point-based SLAM system. A real-time capable version of the system was developed as a final step.
I think this research is very high impact and will be used and built upon for years to come. It’s great to know that the ACRV was one of the first organisations to be publishing in this area!
Can’t wait to see how this work progresses as we start moving closer to high-level understandings of environments