The Design And Implementation Of Visual SLAM System For Indoor Environment

Simultaneous Location and Mapping is the key technology for intelligent robot to realize autonomous navigation.The construction of the SLAM system is generally based on the sensor of the robot itself,this paper mainly studies the SLAM solutions for the indoor environment,the mature solution is to use lidar as the main sensor of the system for environmental information collection.However,lidar has the disadvantages:expensive and bulky,and is not suitable for small,low-cost robot systems.In recent years,with the advent of various low-cost,high-performance vision sensors,the SLAM solutions based on vision sensors have received wide attention from researchers.In these sensors,RGB-D vision sensor is very suitable as the main sensor of SLAM system because it can simultaneously acquire RGB-D images and depth images of environment.Therefore,SLAM based on the RGB-D vision sensor is the most concerned research direction in the robot field.In this paper,we also use RGB-D vision sensor to build a SLAM system that is oriented towards indoor environment,high-precision and real-time scenes.The main research work is as follows:1)We understand the role of the SLAM system in the entire robot system and complete the requirements analysis through studying the modern intelligent robot architecture.According to the requirements analysis,we also complete the design of RGB-D SLAM system framework and module functions.2)We study the imaging model of camera,and use the Kinect camera to realize the data acquisition function.,and also use the camera calibration toolbox of Matlab to complete the calibration experiment of Kinect camera.3)In the front-end visual odometer,firstly,the advantages and disadvantages of the commonly used feature extraction algorithms are compared through experiments.The ORB extraction algorithm is selected as the feature extraction algorithm of the system,and its shortcomings are improved,so that the feature points extracted are distributed as much as possible to the whole image.Then,the Bruce-Force method is used to complete the feature points rough matching,and the RANSAC algorithm is used for optimization.Finally,the RANSAC and ICP algorithm are used to compute the inter-frame camera motion parameters.4)In the back-end,the graph optimization model is used instead of traditional filter model for global pose optimization.Firstly,the key frame selection strategy of the system is determined.Then,based on the selected key frames,the closed-loop detection based on image registration is realized.Finally,the G2 O solver is used to construct and compute the global pose graph,and PCL library is utilized to complete the construction of the global point cloud map.5)Using the TUM dataset from Technische University of Munich and real scenes to test accuracy and effectiveness of the system respectively.The experimental results show that the SLAM system based on RGB-D vision sensor can meet the demands.