Visual Inertial Navigation And Semantic Mapping In Dynamic Scene

In recent years,simultaneous localization and mapping(SLAM)has been widely used in mobile robots,unmanned driving,AR/VR and other fields.Among them,due to the low cost,high popularity and rich information of visual sensor,visual slam has become the current research hotspot.However,the visual sensor has its own defects,resulting in the visual slam effect by the scene.The introduction of inertial measurement unit(IMU)can make up for the scale information and fast motion pose estimation of visual slam,and the visual sensor can also optimize the offset of IMU.At present,visual slam is generally based on the assumption that the environment is static,and the moving objects in the actual scene will inevitably affect the accuracy of the algorithm.In this paper,a slam system is proposed to fuse visual information and inertial information in a dynamic scene.This paper deduces the preintegration of IMU and its error,and analyzes the sensor fusion method and objective function solution method of initialization and back-end optimization.The paper compares the two methods of multi-view geometry and optical flow to deal with dynamic points,and proposes a performance improvement scheme that does not extract feature points in non-key frames and tracks existing feature points with optical flow.The paper combines 2D target detection and point cloud segmentation to construct a semantic map and an octo-map,and simulate local path planning.The slam system of this paper is tested on openloris dataset,and compared with the mainstream visual slam.The method in this paper supplements the scale information of IMU,eliminates the interference of dynamic information,reduces the absolute trajectory error of the system,and effectively improves the stability and robustness of positioning.At the same time,the system constructs a variety of map forms,which can provide support services for navigation and other applications.