Learning scene general feature and specific feature-based six-degree-of-freedom camera pose estimation system

Abstract

The invention discloses a six-degree-of-freedom camera pose estimation system based on universal features and specific features of a learning scene, and belongs to the technical field of visual positioning. The method aims at solving the problems that scene sharing parameters and scene specific parameters in existing visual positioning interfere with each other, and gradient conflicts exist in the training process. Comprises: a basic network module for performing feature extraction on an nth task scene image In collected by a six-degree-of-freedom camera based on shared parameters and specific parameters obtained by training to obtain an image extraction feature Fn; the shared parameter phi and the specific parameter phi n are calculated by adopting an average gradient algorithm; the regression layer is used for predicting an image extraction feature Fn to obtain a 3D scene coordinate Dn and 1D uncertainty Un; and the attitude estimation module is used for calculating a rotation matrix Rn and a translation vector tn of the six-degree-of-freedom camera based on the 3D scene coordinate Dn and the 1D uncertainty Un by adopting a PnP algorithm so as to determine the camera pose Tn in the nth task scene. The method is used for camera pose estimation.

Images