A Robust Visual SLAM via Lightweight Dense Feature Extraction and Dual-Attention Graph Matching Network

With the advancement of computer vision and deep learning, simultaneous localization and mapping (SLAM) has become increasingly important for mobile robot applications. Traditional SLAM relies on handcrafted feature extraction algorithms, which perform poorly in scenarios with significant lighting changes or sparse textures. To address these challenges, this article proposes a robust visual SLAM (LDGM-VSLAM), which consists of lightweight dense feature extraction and a dual-attention graph matching network. First, a lightweight dense feature extraction network is designed for efficient extraction of image keypoints. Second, a graph matching network with a dual-attention mechanism is introduced, aggregating keypoint information via a graph neural network (GNN). Self-attention and cross-attention mechanisms are employed to perform weighted keypoint matching between adjacent image frames. Then, RANSAC and a chi-square test are applied to filter out outliers by validating the statistical consistency of the matched points. Finally, the feature extraction and matching network is integrated with oriented FAST and rotated BRIEF (ORB)-SLAM2, resulting in a visual LDGM-VSLAM. Experimental results show that the proposed method reduces the absolute trajectory error (ATE) by 25.9% compared to SuperPoint-SLAM and shortens the runtime by 10.5 s. In scenarios with sparse textures and significant lighting changes, this method maintains high feature extraction efficiencies and mapping capabilities, providing reliable assurance for the autonomous navigation of mobile robots.