Research On Multi-Sensor Fusion For Indoor Mapping An Localization

With the development of robot technology and the increasing demand from the public,indoor mobile robots with perception and autonomous planning capabilities have become a major research topic.Simultaneous Localization and Mapping(SLAM)is one of the prerequisites for achieving autonomous planning for mobile robots.In complex indoor environments,there are limitations to the information that can be extracted by a single sensor.Therefore,many SLAM algorithms integrate multiple sensors.How to fuse and optimize data obtained from multiple sensors is a major research focus in SLAM.This thesis proposes an indoor mapping and localization algorithm based on multi-sensor fusion from three aspects: feature extraction,loop closure detection,and multi-sensor fusion.The main work is as follows:(1)A deep learning-based feature extraction algorithm is proposed.Traditional feature extraction algorithms are used to enhance the data for local features,thereby improving the algorithm’s generalization ability.At the same time,Super Point and binary layers are combined to speed up feature matching.For global features,a triplet network and Resnet50 are combined to improve the algorithm’s ability to measure image similarity.The method is trained and tested on the public Robot@Home dataset,and the experimental results show that this algorithm is comparable to the ORB algorithm in scene matching accuracy when the pixel distance is less than 2,slightly weaker than the SIFT algorithm;and it is better than the SIFT algorithm and ORB algorithm when the pixel distance is greater than 2.(2)A loop detection algorithm based on RGB images is proposed.The bag-of-words model and global features are combined to perform two-stage matching for loop detection,thereby reducing the false detection rate.Specifically,the algorithm first selects part of the images through local feature filtering,and then selects the most similar images through global features to achieve loop closure detection.The method is tested on the public New College dataset,and the feasibility of the algorithm is verified.(3)A SLAM algorithm based on the fusion of 2D lidar and RGB-D cameras is proposed.In the mapping process,this algorithm uses a 2D lidar to construct a 2D grid map,in which each grid stores the current pose and corresponding global and local descriptors of RGB images.By incorporating visual and lidar information into the map,the robustness of the SLAM algorithm is improved.In the localization stage,the algorithm first selects candidate regions through feature matching results,and then uses 2D lidar information for more precise matching,thereby achieving more accurate localization results.At the same time,the algorithm also uses the ICP algorithm to calculate relative pose transformation,but does not store 3D feature points for a long time.The coordinate information is only used for shortterm pose estimation,which to some extent reduces the storage capacity of the map.