Research On Visual SLAM Algorithm Of Wheeled Robot Based On Multi-feature And Multi-sensor Fusion

With the development of intelligent robot technology,wheeled mobile robots have gradually entered our lives,appearing in shopping malls,stations,hospitals and other places.To improve the autonomous mobility of wheeled robots,it is of great significance to study the SLAM(Simultaneous Localization and Mapping)algorithm for wheeled robots.In recent years,visual SLAM has gradually been applied to wheeled robots due to the advantages of low cost,rich information and easy installation of cameras.However,it suffers from accuracy degradation and tracking failure in low-texture and robot-steering scenes.To solve these problems,this thesis researchs and improves RGB-D SLAM based on point features and monocular SLAM based on multi-sensor fusion.The main work is as follows:(1)To verify the performance of the improved algorithm,a two-wheeled mobile robot platform is designed and built.This platform adopts PC,Raspberry Pi,servo driver three-level control and drive method.It is equipped with a visualinertial camera,encoder and lidar.To reduce the measurement error,the visualinertial camera,the internal parameters of the camera and the external parameters of each sensor are calibrated.The application software of PC and Raspberry Pi is designed and developed based on ROS(Robot Operating System),and the tasks of remote control robot movement,sensor data acquisition and SLAM algorithm operation are completed.(2)To solve the problem of point feature-based RGB-D SLAM tracking failure in low-texture scenes,based on the RGB-D depth camera version of ORBSLAM2,an improved algorithm fusing point,line and plane features is proposed.First,in the visual odometry of ORB-SLAM2,line and plane features are added to construct the objective function together with the original point features,and the nonlinear optimization method is used to minimize the objective function to estimate the current frame pose and improve the tracking performance.Then,in the back-end optimization,the Manhattan Frame is extracted from the line and plane features to construct the objective function together with the point,line and plane features,and the nonlinear optimization method is used to minimize the objective function to optimize the keyframe poses,point,line and plane positions in the local map.It is based on local map optimization and Mixed Manhattan World assumption to reduce accumulated error to improve the accuracy and reconstruction effect of the algorithm.Experiments show that the improved algorithm has better tracking effect in low-texture scenes,and improves the robustness,localization accuracy and reconstruction effect of the algorithm,but the real-time performance is reduced,and suffers from accuracy degradation and tracking failure in more challenging scenes,such as feature sparse and robotsteering scenes.(3)To the problem of improved multi-feature fusion algorithm accuracy degradation and tracking failure in more challenging scenarios such as feature sparse and robot-steering,an improved VINS-Mono algorithm fusing monocular camera,IMU(Inertial Measurement Unit)and encoder is proposed.The encoder measurement error term is added to the objective function of VINS-Mono initialization and back-end optimization.The speed calculated by the encoder data is directly fused to improve the observability of the scale to reduce the accumulated localization error and improve the localization accuracy.Moreover,in order to reduce the influence of wheel slip on the localization accuracy,the slip factor is calculated by IMU gyro data to adaptively adjust the weight of the encoder measurement error term in the objective function and the threshold of its robust kernel function.Experiments show that the improved algorithm retains the real-time performance of VINS-Mono,improves the localization accuracy by an order of magnitude,and has high robustness,which is suitable for feature sparse,robot-steering and 3D scenes.(4)Experiments are carried out in indoor and outdoor scenes to compare the performance of the two improved algorithms applied to two-wheeled mobile robots.The experimental results show that the improved multi-feature fusion algorithm can accurately estimate the trajectory,has a good reconstruction effect in general indoor scenes and is suitable for general indoor scenes.But it suffers from accuracy degradation and tracking failure in feature sparse and outdoor scenes.The improved multi-sensor fusion algorithm achieves high accuracy and robustness in both indoor and outdoor scenes,which is suitable for indoor and outdoor scenes.