Abstract
Structured light profilometry is an established optical technique that measures the 3D shape of an object by projecting fringe patterns (usually lines) onto the object surface and by observing the deformed lines under a fixed angle. Today, state-of-the-art structured light profilometry requires three or more unique recordings to analytically determine the full-field height map of the object. This limits the 3D acquisition speed of the application, complicates the optical setup of the measurement system, and induces motion artifacts in the 3D scans when the target moves between subsequent recordings. In this project, we will train a custom neural network to convert a single deformed structured light pattern directly into its corresponding 3D surface map. By doing so, we will effectively solve the correspondence problem between deformed fringe pattern and 3D map using only a single input image. This will hugely increase the 3D measurement speed in real-time applications, with the frame rate of the camera now being the only limiting factor. In addition, the optical complexity and cost of current state-of-the-art optical scanning systems will be significantly reduced, which will create new possibilities in medical imaging, industrial inspection, machine vision, entertainment, and biometric access security applications. Furthermore, we will build on this new strategy to answer the question of whether neural networks can learn to extract high-resolution and absolute 3D information from a single 2D camera image of an object without using any projected lines, dots, or other fringe patterns – much like humans with monocular vision do. This will omit the need for a projection unit in a 3D scanner altogether and will effectively convert any smartphone camera, endoscope, or smart glasses into quantitative 3D scanning systems. This will result in an entirely new single-shot, ultrafast, and fully scalable 3D depth-sensing technique.
Researcher(s)
Research team(s)
Project type(s)