An Improved Visual SLAM Algorithm Based on Graph Neural Network

Feature extraction and matching are irreplaceable parts of a typical visual simultaneous localization and mapping (VSLAM) algorithm. A variety of different approaches (e.g., ORB, Superpoint, GCNv2, etc.) have been proposed for effective feature extraction and matching. However, as far as we know, such methods still face great challenges in extreme angle and texture sparse scenarios. In this paper, an Improved Visual SLAM Algorithm Based on Graph Neural Network (GNNI-VSLAM), inspired by the strong robustness of the graph neural network in the field of image matching, is proposed to solve the problem of feature point extraction difficulties in sparsely textures scenes and feature point matching difficulties at extreme angles. First, the a priori location estimation feature extraction network is proposed to obtain fast and uniform detection and description of image feature points by a priori location estimation and to build accurate and real feature point information. Second, the feature matching network of the graph attention mechanism is proposed to aggregate feature point information through the neural network of the message passing graph, and then use the self and joint attention mechanism for adjacent image frame weighted feature matching. Then, the feature extraction and neural network are merged with the back-end nonlinear optimization, closed-loop correction and local mapping modules of the ORB-SLAM2 system to propose a complete monocular vision GNNI-VSLAM system. Finally, the proposed algorithm is verified by the public TUM dataset and the experimental results show that the absolute trajectory error of the proposed algorithm is reduced by 29.7 % compared to the GCNv2-SLAM algorithm, which shows a good mapping capabilities.