Reasearch On The Optimization Method Of SLAM System Based On Visual-inertial Fusion

The function of simultaneous localization and mapping is very important for mobile robots working in indoor environment.Accurate simultaneous localization is the basis of mapping and the prerequisite for subsequent realization of mobile robot's perception and interaction to the environment.There were three problems studied in this paper: state estimation of mobile robot with pure visual information,calibration of camera and inertial measurement unit,and the optimization of mobile robot's state estimation with multisensor fusion.The purpose of this paper is to achieve the optimal state estimation of mobile robot in indoor environment by selecting pure visual state estimation method,calibrating sensor and obtaining accurate sensor data,the optimization of state estimation with multi-sensor fusion,so as to obtain higher localization accuracy.Firstly,the relevant mathematical tools and concepts that need to be used in the project was been summarized in this paper.Several visual information detection and matching methods and their applicable conditions were analyzed,and the most suitable visual processing method was explored.As for the errors in visual processing,the feature points caused by mismatches were eliminated by error processing algorithm.Based on the known depth information of binocular vision,a state estimation algorithm based on pure visual information was studied.The elimination method of cumulative error in the closed-loop scene was explored in this paper,and the bag of words model was used to reidentify similar scene,and a strategy for calculating and judging the closed-loop of similarity was put forward,so as to realize the globally consistent closed-loop correction.Secondly,in the case of the sensors' parameters were unknown and the errors were included in the output data,binoucular camera and IMU were both calibrated.The projection model of pinhole camera was studied and the internal parameters of binocular camera were obtained by nonlinear optimization algorithm,and the principle of binocular vision depth measurement was analyzed.The error and characteristic of IMU were analyzed,and the IMU calibration algorithm which included calibrating zero deviation and noise of accelerometer and gyroscope was explored.In order to obtain the spatial pose information between multiple sensors,a combined binocular camera and IMU pose calibration algorithm based on linear transformation method was proposed.In addition,the motion state measured by IMU and the motion state estimated by visual information were fused to optimize the state estimation of mobile robot with redundant information.For the computational repeatability brought by the update of IMU,the algorithm of IMU preintegration in continuous time was explored.To solve the problem of large consumption of computing resources in optimization,the strategy of limiting computation amount with a sliding window was explored,and a set of strict keyframe filtering mechanism in sliding window was proposed.An optimization algorithm of state estimation combining with visual information,IMU information and prior information was determined.At last,the hardware and software platform needed to verify the project were built,the calibration of camera and IMU was completed,and the corresponding sensor parameters were obtained.The algorithm was tested in the indoor environment with mobile robot and binocular camera which embed an IMU as experimental objects.The actual localization accuracy of mobile robot was calculated and obtained.