Visual positioning system and method based on feature noise reduction

Abstract

The invention relates to a visual positioning system and a visual positioning method based on feature noise reduction. The system comprises a feature extraction network for performing feature extraction on an input image to obtain image features; the visual positioning network based on feature noise reduction is used for performing visual positioning based on image features and outputting a pose vector sequence, and the pose vector sequence comprises a current estimated pose and a plurality of estimated poses in the future. Compared with the prior art, the method has the advantages that the current and future real-time motion pose prediction can be given according to the current scene under the off-line condition, the multi-positioning-point prediction is realized, and the considerable generalization ability is also realized under the condition that the motion route is not fixed.

Description

technical field [0001] The present invention relates to a visual positioning system, in particular to a visual positioning system based on feature noise reduction. Background technique [0002] The existing positioning technologies mainly include the following: [0003] 1. Structure-based visual localization techniques: Structure-based pose regressors first extract sparse image features. Inter-frame estimation is then implemented, and then closed-loop detection is performed through matching between feature points, such as scale-invariant feature transform (SIFT)-based, directional FAST-based, and rotation-brief (ORB)-based visual SLAM. SIFT and ORB features are widely used in visual SLAM due to their good robustness, discriminative ability and fast processing speed. [0004] 2. The deep learning model PoseNet is applied to the field of camera repositioning: PoseNet is adapted from GoogLeNet and uses a deep neural network to learn the implicit projection relationship of ima...

AI Technical Summary

Problems solved by technology

[0008] 2. Visual positioning technology based on deep learning Due to the limitations of the existing data set acquisition methods, the pose regressor based on deep learning is prone to overfitting during the training process

[0010] 4. In Visual Synchronous Localization and Mapping (VSLAM), in order to estimate the continuous pose of the camera, structure-based SLAM needs to rely on the collected data sets to reconstruct the 3D environment, and then save the environment map, which will be very time-consuming and labor-intensive. and requires a high initial frame

It is difficult for us to unify different local maps in the same world coordinate system. If there is no GPS correction, we can still only complete the local positioning function

Images