Research On Visual SLAM Method Based On Deep Learning

SLAM is a research hotspot in the field of robots,Visual SLAM technology is the frontier of unknown space exploration,Visual SLAM is affected by dynamic disturbances and light variations,deep learning improves the robustness of the system by extracting deep information from images and fusing the traditional visual SLAM data.The research contents of this paper are as follows:(1)The fusion method of deep learning feature points and traditional visual SLAM algorithm is studied.Firstly,the matching effect of feature points extracted by SuperPoint neural network and ORB feature points is compared,and the matching effect of SuperPoint is better than ORB feature points under complex illumination;Then,Then,the SP-SLAM system is constructed by combining SuperPoint neural network with ORB-SLAM2 algorithm.The system uses SuperPoint neural network to extract image feature points and descriptors,which are used for pose estimation and subsequent functional modules.Finally,a comparative experiment on TUM data set shows that the absolute trajectory error of SP-SLAM is reduced by 16.25% than ORB-SLAM2 in the environment of dynamic object interference.(2)A dense map construction method is designed to eliminate dynamic target point cloud.Add image labels to dynamic objects in keyframes by network segmentation model;Then the keyframe with semantic information is obtained by image mask operation,and the point cloud map is obtained by information extraction;Finally,the dynamic point cloud cluster is eliminated according to the color clustering search method,which improves the construction effect of dense map in dynamic environment,multiple types of output maps are obtained after dense map conversion.(3)In this paper,an experimental platform is built to test and evaluate the SP-SLAM system in a real environment.After the initialization of SP-SLAM system,the environmental feature points could be extracted,while the initialization of ORB-SLAM2 failed due to the feature points could not be extracted.In the positioning experiment,the system error of SP-SLAM is 14.34% lower than ORB SLAM2 under the condition of dynamic object interference,and the positioning accuracy is improved,In addition,the system can build a globally consistent dense map after removing dynamic objects in a dynamic environment.