On-Line dense point cloud generation from monocular images with scale estimation

Abstract

This paper introduces an approach for on-line marker-based three dimensional modeling with scale estimation and heightmap construction from monocular images. The presented system is also capable of an off-line marker-less 3D reconstruction from monocular images with increased detail. This method is designed for the flexible use with an Unmaned Aerial Vehicle (UAV); this means that, despite being tested with a Parrot AR.Drone 1.0, it is easily portable to other more capable UAV models. The followed approach was an adaptation of the patchbased Multiview Stereo (PMVS) algorithm for on line point cloud generation. The system achieved 1.05 processed images per second on average, slightly surpassing the planed objective of 1 processed image per second. The height estimation error ranges between 1-1.5% with a manual marker detection and 4-5% with automatic marker detection, which seems accurate enough for autonomous navigation and path planning. As future work, tests with a better UAV, processing time reduction, marker-less height map construction, autonomous indoor navigation and collaborative on-line 3D modeling are planned.