Dense point cloud map construction based on ORB-SLAM2 and cross-attention network

In the field of 3D vision technology, binocular vision technology, as a representative of low-cost and high-efficiency methods, has gained widespread attention and holds superior application prospects in the industrial sector. The current state-of-the-art binocular sparse visual SLAM systems can accurately and reliably estimate camera trajectories and landmark positions. However, these sparse maps are solely used for localization and cannot be employed for other tasks such as obstacle avoidance or scene understanding. To address this, this paper proposes a framework for constructing dense point cloud maps in binocular sparse visual SLAM systems. This framework enhances the traditional ORB-SLAM2 algorithm by incorporating depth estimation using a cross-attention network to complete depth information and build dense point cloud maps based on its localization capabilities. For the first-frame image, the proportion of incorrectly predicted pixels in the KITTI dataset is only 1.77%. The proposed method is effective and robust on common datasets, suitable for constructing dense point cloud maps in outdoor large-scale scenes. The output model is overall comprehensive and detailed, holding significant practical value in various outdoor scenarios.