Integrated Navigation Algorithm Design And Software Development Based On Visual And Inertial Fusion

With the rapid development of robotics,autonomous driving and other fields,the accuracy and applicability requirements of high-precision navigation positioning and attitude estimation technology are getting higher and higher.Compared with the limitations of GNSS technology in occluded environments and the high cost of the Lidar-SLAM solution,the navigation and positioning solution based on visual and inertial fusion has significant advantages.This article is based on practical engineering applications.In view of the fact that most of the current open source VINS solutions are prone to drift under the interference of dynamic factors such as indoor and outdoor lighting changes and moving pedestrians and vehicles,a stable optimization algorithm is added to the front end to improve the performance of feature extraction,matching and tracking.Accuracy and running speed,and a non-linear optimization framework with a tight combination of vision and inertia is used in the back end to improve the robustness of the navigation system.Integrate the above content,write a set of engineering application software MVINav suitable for low-cost sensors,and reserve interfaces for later fusion of other sensors such as lidar and GNSS.The article specifically made the following research:(1)Two aspects of the loosely coupled initialization process based on visual SFM are optimized.Firstly,the MSAA anti-aliasing technology is used to eliminate obvious aliasing issues such as jagged edges in the image by 2x sampling level,resulting in significant improvement in image quality on sensors with low resolutions.Secondly,to address the shortcomings of traditional optical flow methods,such as easily falling into local optima and slow processing speed,an optimization method based on MipMap optical flow is proposed.By using Gaussian filtering to downsample the original image and generate a series of different scale layers,key points are extracted by the feature extraction algorithm(1)Introducing the observation model of precise point positioning and the error terms involved in the positioning process in detail.According to the least squares and Kalman filter parameter estimation methods,the process of parameter estimation in precise point positioning is deduced.(2)A combination navigation software called MVINav was developed based on the ROS platform using C++ on the Ubuntu system.The software integrates singlecamera visual and inertial navigation by employing a tightly-coupled scheme and a nonlinear optimization strategy similar to those used in OKVIS and VINS＿Mono.The software includes modules for sensor detection,feature extraction,initialization,state estimation,loop detection,and global pose graph optimization,which satisfies the requirement of real-time operation.(3)MVINav was comprehensively tested using the TUM dataset in complex scenarios.The test results showed that the accuracy of MVINav reached 0.03 m in simple indoor scenes with constant illumination,0.5-0.8m in fast moving corridor scenes,0.6m in complex hall scenes with significant changes in illumination,0.4m in pipeline scenes with partial image failure relying on IMU for pose estimation,and 6-15 m in outdoor scenes with significant changes in illumination and complex motion interference.Compared with other open source solutions,MVINav achieved good accuracy performance in some scenarios.Overall,the accuracy of MVINav met the requirements for engineering applications in practical tests.This article contains 158 figures,23 tables,and 101 references.