Research On SLAM Method Of Geometric Constraints Based On Visual-Inertial Fusion

Visual Simultaneous Location And Mapping(VSLAM)is a technique used by mobile robots to collect information about their environment using camera sensors in order to achieve self-localization and construct a map of their surroundings in unknown environments.Research on visual SLAM plays a crucial role in achieving intelligence and automation in mobile robots.However,as the demands on the working environment of mobile robots become increasingly stringent,the information obtained by a single sensor is limited in dynamic or weakly textured environments and cannot meet the requirements for accurate pose estimation.Therefore,multi-sensor fusion positioning technology has emerged as a result,among which the visual-inertial localization and navigation technology using a combination of cameras and Inertial Measurement Unit(IMU)sensors has shown outstanding performance.However,how to accurately obtain usable information in complex dynamic environments is still a pressing problem that needs to be solved.Current Dynamic SLAM algorithms use semantic information segmentation to identify dynamic objects in a scene,but they face problems in recognizing unmarked dynamic objects.To address this issue,a static plane-based approach is used to remove dynamic objects by using a plane segmentation network to extract static planes in dynamic scenes.The improved PlaneRecover plane segmentation network is used to retrieve static planes and plane parameters,and a plane loss function is constructed to integrate and refine plane segmentation details.The plane parameters are then used to solve for 3D point information,avoiding errors in system pose estimation caused by unknown dynamic objects.To address the issues of time-consuming and feature clustering in point feature extraction,an improved AGAST feature point uniformization extraction algorithm is employed.Feature points are extracted from the constructed image pyramid layer by layer,with each layer’s root node divided into a grid.The grid is refined to retain only the feature point with the highest response value,reducing feature point clustering and improving inter-frame pose estimation accuracy.To address the issue of feature matching errors in mobile robots when dealing with weak textures and large pose variations,an improved dual-feature mis-matching method based on point-plane feature geometric constraints is used.In the plane feature matching step,physical plane parameters are used for feature matching,and a mechanism for eliminating errors in plane feature matching is designed based on the principle that the angle between plane normals is invariant.This results in correct plane feature matching and the calculation of point feature information with planes.In the point feature matching step,point feature matching with different plane information is directly eliminated,and a mechanism for eliminating mis-matching of point features within the same plane is designed based on the three-focal tensor and RANSAC principles.By determining the positions of point features within the same plane in consecutive frames,the threshold of the point feature tensor is calculated to distinguish and retain the correct point feature matches.To address the issue of large pose estimation errors caused by poor image quality in complex environments when using a single visual sensor,the camera and IMU are combined to improve the robustness and mapping accuracy of the SLAM system by utilizing their complementary nature.To verify the feasibility of the proposed algorithm,we tested the point-plane feature-based SLAM algorithm and the visual-inertial fusion positioning algorithm based on dual geometric feature constraints on the TUM and OpenLORIS-Scene public datasets.We also ran real-world scenario datasets on our hardware platform to validate our proposed algorithm.The results showed that our method effectively improved feature matching accuracy,localization accuracy,and system robustness in indoor environments.