End-to-end visual positioning method and system

Abstract

The invention relates to an end-to-end visual positioning method and system, and the method comprises the steps: obtaining a training data set which comprises a plurality of frames of continuous source images,establishing a pose regression network model according to each continuous source image,the method specifically comprises the following steps: for each source image, predicting a correspondingdepth map according to the source image based on a deep convolutional neural network, determining a composite image through a back projection method according to the internal parameters and the depthmap of the camera, determining a pose regression network model according to each source image and the corresponding composite image, and based on the pose regression network model, obtaining an absolute pose of the to-be-measured image according to the to-be-measured image. The method is based on a deep convolutional neural network Depth CNN, a corresponding depth map is predicted according to asource image, and a synthetic image is determined through a back projection method, thus, the pose regression network model is determined, end-to-end visual positioning is realized, the absolute poseof the to-be-measured image can be accurately determined, and the positioning precision is improved.

Description

technical field [0001] The present invention relates to the field of computer vision technology and SLAM (Simultaneous localization and mapping, synchronous positioning and mapping), in particular to an end-to-end visual positioning method and system based on an online geometric data augmentation strategy. Background technique [0002] Visual localization is an important link in mobile robots, automatic driving and augmented reality, which refers to estimating the pose of a camera through images. [0003] The current mainstream visual localization algorithm is based on geometry. Given a captured image, it is first necessary to use the SFM (structure-from-motion) algorithm to perform 3D reconstruction of the scene, and the reconstructed 3D model points will be assigned one or more feature descriptors. When a query image is given, the feature points are first extracted and the descriptor is calculated, and then the 3D point that is most similar to the feature point is searche...

AI Technical Summary

Problems solved by technology

Although the end-to-end method can overcome some shortcomings of the geometric method, its localization accuracy is limited by the sparsity of the training data

Because the acquisition of positioning data is relatively difficult, the poses in the training database of the end-to-end model usually only contain a small part of the positioning space, which makes the network very easy to overfit during the training process.

Most of the early work focused on designing new network structures or loss functions to improve the generalization ability of the network, but the accuracy improvement is not obvious

Images