Constructing a three-dimensional point cloud

Abstract

A method of constructing a 3D point cloud, the method comprising: receiving a hyperspectral image (34, 36) from each of two hyperspectral cameras (16, 18) mounted onto an entity, determining positions of regions 54 in the images, determining 3D positions of each region based on the positions of the regions in the images and the relative distance and orientations of the cameras, constructing a 3D point cloud including a point at each determined 3D position. The position of the image regions may be determined by finding a one-to-one best match between the hyperspectral signatures of pixels in the two images and identifying matching pixels as the regions. Alternatively the position of the regions may be determined by clustering pixels in the two images into groups and identifying each group as a region. The one-to-one match and / or cluster groupings may be determined using geometric rules, a rule-based algorithm, a neural network, or an unsupervised machine learning model. The regions may be input into a trained classifier.

Images