| Simultaneous Localization and Mapping(SLAM)technology is one of the core technologies for autonomous navigation of mobile robots and has attracted much attention.Monocular SLAM has become the main research object of many researchers because of its simple structure,low cost and scalability.However,traditional monocular SLAM still has many problems in practical applications,especially in feature point matching and dynamic environment.During the matching process,the camera is easily affected by rotation and viewpoint changes,and when performing loopback correction,it is difficult to perform effective loopback verification of dynamic detection targets due to insufficient words in the bag-of-words model.This all leads to a reduction in the accuracy of camera pose estimation,which can directly affect the performance of back-end localization and map building.Therefore,it is necessary to introduce deep learning techniques in monocular SLAM to obtain better matching ability and scene perception.In this paper,a Graph Neural Network(GNN)is constructed using deep learning-based principles and replaced with the feature matching module of the VSLAM system.The feature extraction of ORB is used to combine the two and incorporated into the visual odometer to improve the matching accuracy and camera position estimation accuracy of the visual odometer.A combination of improved NBSCAN clustering algorithm and K-means algorithm is used to construct the bag-of-words model in the back-end loopback detection.The improved algorithm is more accurate in handling non-convex and dense data sets,increases the bag-of-words word size,and improves the system’s ability to cope with dynamic environments.The main works are as follows.(1)To address the traditional feature point matching which is easily affected by rotation and perspective change,this paper adopts a deep learning approach to solve the feature matching problem after feature point detection.Firstly,the ORB extractor is used for feature point acquisition,and then the Super Glue network model is used for feature point matching to improve the accuracy of feature matching and achieve the purpose of optimizing the camera pose.(2)In order to cope with the dynamic detection target in the experimental scenario,this paper optimizes the traditional bag-of-words model.The improved dictionary not only adds the outlier description points to the dictionary,but also the outlier words have a larger proportion in the dictionary,which expands the scope of the dictionary and makes the loopback detection more accurate and improves the system’s ability to cope with the dynamic environment.(3)The indoor oilfield inspection robot is selected as the research object.The improved algorithm is fused into the SLAM system as a whole,applied to the inspection robot,and tested in the indoor environment.The results prove that the algorithm in this paper is better than the traditional ORB-SLAM2 algorithm in terms of localization accuracy.. |
