Research On SLAM Algorithm Based On Multi-sensor Fusion

In recent years,the Simultaneous Localization and Mapping(SLAM)technology represented by vision has been greatly developed in autonomous flying robots,logistics distribution robots,cleaning robots and autonomous driving.However,the current technology still can't be solved in the face of complex environments.The main problem faced by the visual inertial navigation system in the large outdoor scene is that the space complexity and time complexity of the system will increase rapidly when there is a loop,which leads to the inability to locate in the real time.The long term system operation without loop will produce cumulative errors.However,GPS as a global sensor can achieve drift-free localization in large-scale environments.It can be seen that GPS and visual inertial navigation systems have certain complementary characteristics.Therefore,this article uses cameras,inertial measurement unit and GPS sensors to implement a multi-sensor-based fusion localization and mapping system.The main contents are as follows:(1)The sensors used are mainly low-cost sensors.In order to provide good values for the fusion system,this paper carries out sensor calibration and data preprocessing,including monocular camera calibration,IMU internal parameter calibration,and camera and IMU joint external parameters calibration and GPS data preprocessing.GPS data processing includes strictly limiting the threshold of the received satellites and covariance matrix and converting GPS data to the East North Up(ENU)global coordinate system.(2)In order to improve the accuracy of the VI-SLAM system,this paper uses the fourth-order Runge-Kutta integration to integrate the IMU and prop oses a tracking outliers elimination scheme that combines inverse optical flow and ORB descriptor technology.Experimental results shows that the accuracy of the fourth-order Runge-Kutta integration is higher than that of the median integration.The optical flow tracking results are further purified,the system has higher accuracy.(3)Aiming at the problems of VI-SLAM system in large outdoor scenes,this paper proposes a GPS and VI-SLAM tightly coupled algorithm based on nonlinear optimization,which estimates transformation matrix between the GPS position in the global coordinate system and VI-SLAM online during initialization The system uses the IMU residual to derive the global position residual to construct th e tight coupling between the IMU and the camera and the IMU and GPS.The GPS position residual is added to the optimization objective function of the system,which fully considered the sensor data association.In addition,in order to compare with the traditional localization method,the IMU and GPS fusion based on the error state Kalman filter is realized.Finally,the proposed algorithm is verified by datasets and real scenes.The results show that the root mean square error of the improved VI-SLAM algorithm in most sequences of datasets is lower than that of the current open source algorithm and the robustness is better in the actual scene.Compared with the open source VI-SLAM algorithm,the proposed multi-sensor fusion algorithm eliminates the cumulative error and the localization accuracy is better than the traditional localization method.