Pose calculation method and system combining deep learning and geometric optimization

Abstract

The invention discloses a pose calculation method combining deep learning and geometric optimization. The system comprises a pose estimation module PoseNet and a depth estimation module DepthNet, a pose estimation method based on deep learning and an optimization strategy based on geometric constraints are integrated into a visual odometer frame, luminosity errors, feature point reprojection errors, adjacent inter-frame constraints and constraints in a local map formed by a section of continuous frames are calculated, and real-time and accurate pose estimation can be carried out.

Description

technical field [0001] The invention belongs to the field of visual odometry in the field of computer vision, in particular to a pose calculation method and system combining deep learning and geometric optimization. Background technique [0002] Vision-based pose estimation (Visual Odometry, hereinafter referred to as VO) enables the robot to accurately locate in an unknown environment only relying on the data collected by the camera. In the past ten years, the VO framework based on the feature point method and the photometric error method has made great progress. The traditional VO based on the geometric method can perform very Lubang positioning in most cases, but in the absence of feature points or When the camera exposure is unstable, the accuracy of pose estimation will drop or even fail. In recent years, due to the development of deep learning and the explosive growth of data volume, learning-based VO has attracted more and more attention. It has two main advantages:...

AI Technical Summary

Problems solved by technology

[0004] At present, this type of method can obtain excellent accuracy on the monocular depth estimation task with the help of binocular video training data, but it lags far behind the geometric method-based VO on the monocular pose estimation task.

The reason is that the interpretability and generalization ability of the neural network on the 3D geometric tasks that are strictly proved by mathematics is currently weak; this type of method can only estimate the pose change between two frames, ignoring the time sequence between multiple frames. Dependencies and geometric constraints on the local map, after a long period of error accumulation, the result of the pose estimation will deviate seriously from the true value

Images