Abstract: Aiming at the problem of low positioning accuracy of visual simultaneous localization and mapping (SLAM) based on point features in indoor environments with weak textures, a point-line-plane multi-feature visual SLAM method based on Atlanta constraints is proposed. To address the issue of decreased recognition accuracy due to the increase in the number of Manhattan coordinate systems, an Atlanta world detection method based on semantic information and gravity direction is employed. To overcome the challenge of insufficient accuracy in traditional plane detection algorithms, an improved yolov8 semantic plane detection method is utilized. To solve the potential cumulative error problem arising from the fusion of Atlanta SLAM with traditional point-line-plane SLAM, a mutual correction method between Atlanta coordinate systems is adopted to eliminate the cumulative error impact of point-line-plane feature positioning on the Atlanta coordinate system. In Atlanta scenes, the corrected Atlanta coordinate system is prioritized, achieving drift-free rotational estimation through the matching of the current plane with the map; in non-Atlanta scenes, rotational and translational estimation is achieved by matching point-line features between adjacent frames. Experimental results show that the improved yolov8 semantic segmentation network has a 15.5% increase in plane segmentation mAP value compared to yolov8; the average absolute trajectory error is reduced by 29.3% compared to Manhattan SLAM.