Local feature cameras for structure from motion (SFM) problems with generalized cameras

摘要

Methods and systems for extraction of 3D geometry from a plurality of generalized camera images by a device that comprises an electronic circuit are provided. Methods include identifying an x and y coordinate, an orientation, and a scale for each of one or more feature locations in each of the generalized camera images, extracting a local image feature at each feature location, generating a feature camera centered on each feature location, identifying groups of feature cameras providing consistent triangulation opportunity, and triangulating each identified feature camera group by finding the 3D point that minimizes an error term.

主权项

1. A method for extraction of 3D geometry from a plurality of generalized camera images by a device that comprises an electronic circuit, the method comprising:
identifying, by the electronic circuit, one or more feature locations in each of the plurality of generalized camera images, each feature location comprising an x and y coordinate, an orientation, and a scale; extracting, by the electronic circuit, a local image feature at each feature location; generating, by the electronic circuit, a virtual feature camera centered on each feature location; identifying, by the electronic circuit, groups of feature cameras, each group of feature cameras comprising one or more of the virtual feature cameras, wherein each group of feature cameras is identified as providing consistent triangulation opportunity by determining virtual feature cameras that provide real-world coordinates for a same local image feature that are within a threshold value; and, triangulating, by the electronic circuit, each identified feature camera group by finding the 3D point that minimizes an error term, wherein the generating the virtual feature cameras comprises:
obtaining, by the electronic circuit, a feature camera center pixel from a camera pose at a first time; andcreating, by the electronic circuit, virtual linear camera models at each feature location of each extracted local image feature separately, each virtual linear camera model being centered on its respective feature location based on the obtaining.