Visual SLAM Algorithm Based on Deep Learning Feature Extraction and Mismatch Elimination

To address the problem of poor system robustness of visual SLAM under sparse texture environment and fast camera rotation conditions, this paper improves the front-end based on the ORB-SLAM2 framework. The front end uses the Geometric Correspondence Network v2 (GCNv2) neural network based on deep learning to extract feature points. Due to the local characteristics of image features, there are mismatches of feature points. Fast approximate nearest neighbor algorithm (FLANN) and progressive sampling consistency algorithm (PROSAC) are used for feature matching and mismatch removal. Tested on the TUM public dataset, the experimental results show that the improved algorithm has no tracking loss under sparse texture environment and fast camera rotation conditions, more uniform distribution of extracted feature points, higher efficiency of false match rejection, smaller absolute trajectory error of the system, higher localization accuracy, and better robustness.