Research And Implementation Of SLAM Algorithm Based On RGB-D Vision And Inertial Fusion

In recent years,a variety of mobile robots have appeared in front of people.They can facilitate people's lives and can replace people to complete some high-risk tasks.One of the core algorithms of this robot is the Simultaneous Localization and Mapping(SLAM)technology based on vision.Since the RGB-D camera can directly measure the depth information of the image pixels,the depth calculation process of the visual SLAM becomes simple,and it can be directly used for the construction of dense point cloud maps in the indoor environment,which can effectively improve the accuracy and robustness of the visual SLAM system.However,due to the defects of the vision sensor itself,it is unable to effectively extract image features under weak texture and fast motion,resulting in loss of tracking.The Inertial Measurement Unit(IMU)has a good estimate under fast motion,which can make up for the lack of visual sensors.Therefore,based on the method of RGB-D vision and inertial fusion,this thesis studies the SLAM technology of indoor mobile robots,and proposes a high-precision and robust real-time visual-inertial SLAM system.The specific research content is as follows:This thesis first summarizes the coordinate system involved in the thesis and the transformation form of the robot's pose coordinate.The pinhole camera model,IMU measurement model and kinematics model are analyzed,and internal and external parameters of the camera and IMU sensor are calibrated.In view of the fact that ORB feature points have a small number of feature detections under weak textures,and feature clusters are prone to appear in texture-rich scenes,the image pyramid is built to improve the tracking robustness under weak textures,and the quadtree algorithm is used to filter at the same time.Feature points reduce repeated texture information and improve image tracking efficiency.In addition,in order to reduce the mismatch rate of image features,a bag of words model for training image feature descriptors using the K-means++ algorithm is proposed,and the word matching method is used for processing,reducing candidate matching pairs,thereby improving the accuracy and speed of image feature matching.The SLAM algorithm of vision and IMU fusion is deeply studied,and an IMU initialization method is proposed.Taking pure visual tracking's pose estimation and IMU pre-integration results as a priori conditions,the IMU's gravity direction,velocity and bias are optimized and solved,and then the overall optimization is performed,and accurate initialization results are obtained.In the process of feature tracking,the EPn P algorithm is used to solve the pose,and then the IMU pre-integration data is merged,and the graph optimization method is used for tight coupling optimization,which improves the tracking accuracy of the system.When the image tracking is lost,the tracking state can be quickly retrieved through word matching between key frames,thereby improving the tracking robustness of the system.Aiming at the problem of RGB-D camera with depth measurement noise,a filtering method is proposed to eliminate points with large measurement noise,which improves the accuracy of map reconstruction.In view of the large amount of point cloud data in large scenes,the point cloud is divided into three-dimensional grids to eliminate repeated texture information,thereby greatly reducing the storage space of the point cloud map.Finally,in order to verify the accuracy and real-time performance of the system,this thesis compares and analyzes the Eu Ro C indoor datasets with other mainstream open source SLAM systems.The experimental results show that the method in this thesis is better than other methods in different difficult scenarios,and the root mean square error is about 6cm.At the same time,the calculation speed is about 20 Hz,which can meet the real-time requirements of the system.