Research On Visual-IMU Tightly Coupled Indoor Location Method Based On Scene Feedback

In this era of increasing artificial intelligence,the rapid rise of high-tech and highintelligence industries such as autonomous driving and autonomous positioning of drones,has given rise to high demand for indoor positioning.The traditional method of positioning,which is based on GNSS and IMU is highly prone to failure in a GNSS or weak GNSS indoor environment.Thus,how to solve the problem of autonomous positioning in indoor environments without GNSS is the core of autonomous driving and intelligent robots.The visual SLAM method can attain the positioning of an unknown environment without relying on the a priori information of that environment and the visual sensor.Furthermore,considering the IMU and the visual SLAM method have very good complementarities,positioning based on the fusion of the IMU and visual SLAM as IMU-visual SLAM has become a popular research in indoor positioning.The traditional visual SLAM method based on single-point features has low robustness due to lack of features in the face of low-texture scenes,but line features are very rich in scenes such as white walls.Thus,this article first establishes based on point-line features an Inter frame tracking framework improve system robustness where different scenes will lead to differences in the performance of point and line features.Therefore,this paper proposes to build an adaptive estimation model of point and line features based on scene segmentation.Finally,after establishing an adaptive point-line weighted tracking model,a system initialization model based on the IMU preintegration theory and a tight coupling method between the IMU and the visual end are established,thus establishing a vision-IMU tightly coupled monocular vision SLAM indoor positioning based on scene feedback system.The specific contents of this article are as follows:(1)The front end establishes a multi-feature observation model: First,fully analyze the scene with a less robust single point feature,compare and analyze the advantages of line features,complete the interframe tracking of point lines,and establish the multiple feature positions Pose estimation model.(2)Considering the different scenes,the performance of point-line features will produce significant differences,so this paper segmented the scene based on the U-net network taking into account the different image positions which will cause distortion to the image block weight.Finally,the weighting ratio of the entire image is established according to distance weighting,and then the Schubauer equation of the adaptive estimation model at the visual end is established.(3)According to the IMU pre-integration theory,the relationship between the IMU measurement value and the motion estimation is derived to complete the system initialization.The tight coupling of IMU and vision establishes the IMU-visual tight coupling pose estimation based on Levenberg-Marquard method.Finally,the closed-loop detection and global optimization of the system are introduced to complete the scene feedback monocular SLAM system with IMU fusion.Through the research content and method in this paper,the visual-IMU tightly coupled SLAM based on scene feedback is realized positioning system.Finally,based on the open source Dataset EuRoc and Trifo Ironsides Dataset,the method proposed in this paper is verified.The results show that compared with the reference method,the method proposed in this paper can further improve the positioning accuracy.