Research On Algorithm Of SLAM Based On Visual-inertial Fusion

Simultaneous Localization and Mapping technology is one of the prerequisites and core technologies for mobile robots to achieve autonomous navigation.Solution based on multi-sensor fusion has always been studied to improve the accuracy and robustness of SLAM.Among which the monocular visual-inertial fusion SLAM system has attracted more attention in recent years.The monocular vision sensor has the advantages of lower price,higher resolution,with no cumulative error.However,when it encounters a weak texture environment,occlusion,fast motion,etc.,positioning cannot be achieved.Also,monocular vision lacks absolute scale information.Inertial Measurement Unit has the advantage of accurate positioning in a short period of time,but it cannot be located for a long time.The fusion of these two can make of their advantages to build a more accurate and reliable SLAM system.This paper mainly studies the SLAM related algorithm of monocular visual-inertial fusion.The main work includes:(1)Study the joint initialization stage of monocular visual-inertial fusion SLAM based on loose coupling.First,we search for suitable algorithm for extracting and tracking the visual front-end feature point.The outliers that appear in the tracking process will be eliminated using the Random Sample Consensus algorithm.Second,study and deduce the IMU pre-integration algorithm to obtain pose estimation of the current moment and the state quantities needed for the back-end optimization.Last,IMU-fusion initialization algorithm based on model selection is proposed,so that the system can automatically select a suitable model according to different scenarios to complete the joint initialization,obtaining the initial value of the system,and improve the robustness of the initialization stage.(2)Study the active failure detection and recovery module of SLAM system.First,on the basis of completing the joint initialization to obtain the initial value of the system,monocular visual-inertial fusion optimization model based on tight coupling is constructed.Second,loop closure detection algorithm based on the bag of words and global optimization algorithm are used to reduce the cumulative error of the system and obtain a globally consistent trajectory estimate.Last,regarding the problem that SLAM systems are prone to failing in special environments,an active failure detection and system recovery module is proposed.This module can automatically detect system failures and restore the system to initialization stage to establish a new independent pose graph,making the system run normally again,and improve the practicality of the system.(3)The experimental platform required to verify the SLAM algorithm in this paper is built.First,calibration experiments are performed on the IMU and the camera to obtain the sensor parameters.Then the SLAM algorithm proposed in this paper is experimentally verified in the Euroc public dataset and in real scene.The experimental results show that the system based on the SLAM algorithm of monocular visual-inertial fusion proposed in this paper has good accuracy,robustness and practicability.