Visual-Inertial Integrated Navigation Based On Deep Learning

With the continuous development of autonomous driving and intelligent robots,autonomous positioning technology has been extensively researched and applied.Visual Inertial Odometry(VIO)algorithm is one of the common technologies of autonomous positioning technology.The VIO algorithm based on deep learning shows excellent competitiveness in robustness,but the existing VIO based on supervised learning takes a lot of time to balance the training of displacement and rotation,and only considers the balance between displacement and rotation,which is too ideal;on the other hand,the network architecture of the current deep learning VIO algorithm is too large and its generalization ability is too weak.In response to the above problems,this article has conducted an in-depth study on the deep learning VIO algorithm.The specific work is as follows:(1)In the existing VIO algorithms based on supervised learning,the magnitude of the displacement label and the rotation label are very different,and the network needs to be trained multiple times to select the appropriate balance coefficient to balance the learning of displacement and rotation.The operation is too cumbersome and it did not consider the order of magnitude balance between the internal displacement and the internal rotation,resulting in insufficient training of the VIO algorithm.To solve this problem,this paper studies the deep learning VIO algorithm based on tag standardization,and converts all displacement and rotation labels in the training set to standard normal distributions,avoiding the need for multiple training to select the balance coefficient,and taking into account the internal displacement and rotation internal order of magnitude balance.This paper is tested on the KITTI data set,and the results prove that the algorithm in this paper improves the displacement accuracy by 42.7% and the rotation accuracy by 68.3% compared with similar algorithms.At the same time,the algorithm in this paper is more robust to data than similar algorithms.(2)On the basis of convenient to obtain displacement and rotation balance through tag standardization,the existing deep learning VIO algorithm is bulky,not lightweight enough,and lacks robustness to image data and inertial data under high data damage rates.In this paper,Liquid Time-Constant Recurrent Neural Networks(LTC-RNN)are studied,and LTC-RNN is used for inertial feature extraction and pose regression,so as to reduce the neural network parameters of the deep learning VIO algorithm and improve the robustness against data damage.This article is trained and tested on the KITTI data set.The results prove that the algorithm in this paper reduces the number of parameters by 55% compared with the same type of algorithm,and at the same time increases the displacement accuracy and rotation accuracy by 1.5% and15% respectively.And it still has good pose estimation performance under high data corruption rate.