Real-time 3D SLAM reaches millimetre-level accuracy

July 08, 2016 //By Julien Happich
French startup Pixmap is officially launching what it describes as a revolutionary 3D real-time robotics localization and mapping technology, Reality Capture.

The company's embedded solution is based on over 11 years of academic research undertaken at the I3S/CNRS (Centre National de Recherche Scientifique) laboratory by the company's CTO, Maxime Meilland together with Andrew Comport, Pixmap's Chief Scientific Officer, a researcher at INRIA Sophia Antipolis where he pioneered work on dense localisation and mapping algorithms.

Both founded the company early 2015 with robotics expert Benoit Morisset (the startup's CEO) shortly after their paper "On unifying key-frame and voxel-based dense visual SLAM at large scales" received the best scientific publication award at IROS 2013 (International Conference on Intelligent Robots and Systems).

The work detailed in that paper stemmed from French DGA's Fraudo project (FRanchissement AUtomatique D'Obstacles or automated obstruction clearance), requiring dense localisation and mapping techniques for robots to traverse uneven ground and surfaces autonomously.

Available under licensing agreements, the algorithms at play unify two approaches commonly used to define dense models, volumetric 3D modelling (using voxel grids) and image-based key-frame representations, into a compact low-memory bandwidth solution supporting refresh rates up to 2kHz.

Dense reconstruction of an entire floor obtained in real-time from a 100 meters trajectory containing 67 key-frames: (top) bird eye view, (bottom) side view of the reconstruction. Illustration taken from Pixmap's IROS 2013 paper.

Reality Capture enables robots and drones to robustly map their environment in 3D and in a photorealistic manner to know their position within the world with a millimetric accuracy. Because the 3D maps are metrically accurate and provide a photo-realistic rendering of the environment, they can be used not only by the robot for path optimization and collision avoidance, but also by remote operators willing to visually revisit sites mapped by a robot.