A visual SLAM method based on double attention feature point extraction network

Abstract

Deep learning-based feature point extraction networks can stably extract feature points in complex environments. however, the extracted feature points tend to be distributed in blank areas. To address this issue, a dual-attention-based feature point extraction network called PCA-GCNv2 is proposed. First, the Feature Pyramid Networks is used for multi-scale feature map fusion, then the dependence relationship between location and channel dimension is modeled by dual attention module, and finally the multi-scale feature map is weighted. The proposed method reduces the number of feature points extracted by the feature point extraction network in the blank region, and improves the accuracy and stability of the extracted feature points. Building on this, the paper further proposes a visual SLAM method based on PCA-GCNv2, called PCAG-SLAM, which integrates PCA-GCNv2 to replace the feature point extraction module in ORB-SLAM2, thus demonstrating the network’s stability in a SLAM system. Simulation experiments using the TUM data set validate the method. The results show that the proposed PCA-GCNv2 feature point extraction network and the PCAG-SLAM method significantly improve feature point extraction accuracy and reduce pose estimation errors compared to other existing methods.