Feature-free photogrammetric 3D imaging with cameras under unconstrained motion

Unmet Need

With the increasing development of camera-based sensors for applications such as virtual and augmented reality, there is more demand for technologies that can image 3D structures and create interpretable digital representations – a practice known as photogrammetry. Photogrammetry is employable for imaging 3D objects of any size even up to topographical mapping of the earth. However, efforts to create reliable photogrammetric imaging at the smaller millimeter-to-micron scale are much less developed. There is a need for simple yet effective photogrammetry methods that are capable of 3D imaging in the mesoscopic (µm to mm) range with accurate representation.

Technology

Duke inventors have developed a software-based method for photogrammetric 3D imaging of mesoscopic-sized objects using completely unconstrained cameras. This is intended to be offered direct to consumers for small-scale imaging applications with simple-to-use, affordably available cameras such as freehand-controlled smartphones. Specifically, the user takes multiple photos of a 3D object using a freely positioned camera at close range to the object. By recording the dynamic 6D positioning of the camera and the pixel-intensity for each photo, the software can provide a reconstructed image with a height map of the observed 3D object. The software is also able to account for camera distortion and reconstruction artifacts through modeling methods that remove such errors. This has been demonstrated to produce accurate height-mapped 3D images of multiple 3D objects (e.g. circuit board and brushstrokes on a painting) with tens-of-micron resolution using sequences of images obtained from an unsupported smartphone camera.

Advantages

• Utilizable with affordable and accessible cameras (e.g. smartphones)
• Accurate in the dimension of tens-of-microns
• Able to estimate and remove distortion effects (e.g. camera motion)
• Removes artifacts of digital reconstruction using machine learning