A method of generating a three dimensional representation of an environment or system

摘要

Method of generating a 3D representation of an environment or system, comprising: processing a representation (eg. images, point clouds) of the environment/system to generate a set of voxels; dividing the voxels into subsets; regularising at least one of the subsets; generating a 3D representation of the environment/system from the voxels thus processed. The environment representation to be processed may be a depth map. Subsets may be formed based on: whether a voxel is observable; presence/absence of data; ability to interpolate data; type of sensor that captures the pictures/depth maps of the real scene; colour/texture/reflectance information; geometrical assumptions; labels/metadata; image recognition data; scene segmentation heuristics (eg. context information). Physical systems modelled may include gas models or rocket plumes. Depth maps may be fused using regularised energy minimisation (eg. L1 or L2 normalisation) acting on Truncated Signed Distance Function (TSDF). Weighting may be used. Noisy data may be smoothed. Unobserved surfaces may be interpolated. Vehicles may be located in the data. Voxel subsets may represent specific surfaces. Regularisation may ensure that surfaces are coherent or smooth and do not erroneously include voxels that are not connected with the surface in the real scene. The regularised voxels are used to create a 3D model.