Research On Mobile Robot SLAM System Based On Deep Learning

In recent years,with the advancement of big data,artificial intelligence,5G and other emerging technologies,the field of robotics has also developed vigorously.Simultaneous Localization And Mapping(SLAM)as the core technology of the robot,can make robots more intelligent and autonomous,and bring more convenience and efficiency to human life and work.Therefore,it has received extensive attention.Visual SLAM has always been a popular research direction in the field of SLAM.Visual sensors have the advantages of low cost,high resolution,rich information,small size,ability to obtain dynamic information,and applicable to various scenarios.However,visual sensors also have disadvantages such as sensitivity to illumination changes,occlusion,and motion blur.Deep learning has more obvious advantages in terms of semantic understanding,multi-modal data processing,adaptability,robustness,and high precision.Therefore,introducing deep learning into SLAM can help make up for the shortcomings of pure visual SLAM.This thesis conducts related research on the direction of using deep learning to replace the pure visual SLAM module.The main work is as follows:(1)Use the SuperPoint network to replace the traditional feature point extraction algorithm.The SuperPoint network is a deep learning neural network that can be used to extract keypoints and descriptors in images.In this thesis,the SuperPoint network is used to replace the traditional feature point extraction algorithm to solve the problem of poor robustness of the traditional algorithm in extreme scenes such as dark night,low texture,and reflection.This thesis uses the SuperPoint network to build a front-end visual odometer,and compares the corner repeatability of SuperPoint and traditional algorithms through experiments.(2)Use the SuperGlue network to replace the traditional feature point matching algorithm.SuperGlue is a deep learning neural network that can be used for feature matching.This article first introduces its operating principle.Then combine SuperGlue and SuperPoint to build a front-end visual odometer based on SuperGlue+SuperGlue(SP).And compare the pose estimation error of SP visual odometry with traditional ORB and SIFT visual odometry through experiments.(3)Use the Yolo-Fastest V2 network to remove dynamic target feature points.Due to the existence of dynamic objects,traditional algorithms are easily disturbed.This article introduces how to use the Yolo-Fastest V2 network to eliminate dynamic target feature points,combine the Yolo-Fastest V2 network with the SP visual odometer,and build a front-end visual odometer based on Yolo+SP(YSP).(4)Design a complete SLAM system YSP-SLAM based on deep learning front-end visual odometry.In order to further compare the performance of the algorithm,this thesis builds a complete SLAM system YSP-SLAM on the basis of the YSP front-end,and conducts a comparative experiment with the traditional ORB-SLAM2.In order to verify the performance of the algorithm in real scenes,this thesis independently designed and built a mobile robot software and hardware platform according to the experimental requirements,and conducted comparative experiments on extreme scenes such as low texture,high reflection,and dynamics where traditional visual SLAM performance is poor.