Vision localization method for intelligent vehicles in low-texture environments

In order to address the problem that the intelligent vehicle will generate localization error due to the inability to extract enough feature points in the image under the low-texture environment, a detector-free method is proposed to accurately solve the intelligent vehicle position pose. Firstly, we perform dense matching at pixel level based on the improved LoFTR method for low-texture surfaces; secondly, we use DBSCAN to cluster the matching and select regions with higher matching confidence; finally, we use RANSAC to filter the outliers within the clustered regions to obtain accurate positional results. The experimental results on the public dataset show that the proposed method has good localization results.