一种融合深度学习双目深度估计的视觉SLAM系统

he primary bottleneck in current visual SLAM (Simultaneous Localization and Mapping) systems is the inability to obtain accurate depth images in large-scale outdoor scenarios. Monocular depth estimation methods based on deep learning fail to capture scale information and often lack detail in complex scenes. This paper proposes an improved real-time inferable binocular depth estimation network and integrates it into visual SLAM, effectively creating an RGB-D-like SLAM approach. This method is capable of real-time operation on the KITTI dataset and constructing dense point cloud maps. A performance evaluation and comparison of the proposed method with other mainstream SLAM techniques on the KITTI dataset demonstrate that it achieves competitive localization performance compared to traditional binocular visual SLAM systems.