Level-LOC: Loop-Optimized GNSS-Visual-Inertial SLAM using Multi-Height Semantic Building Contours

This paper presents Level-LOC, a Loop-Optimized GNSS-Visual-Inertial SLAM Using Multi-Height Semantic Building Contour features to improve loop closure robustness in large-scale urban environments. Building and ground regions are segmented using YOLOv11, and stereo depth is fused to generate a pseudo-LiDAR point cloud. Static contour points are sampled at fixed height intervals from 1 m to 5 m above the estimated ground plane and projected into birds-eye-view contour layers. During loop closure, a contour-based geometric verification step filters DBoW3 visual loop candidates, rejecting false positives caused by dynamic objects and illumination changes. On a 1 km handheld urban dataset, the proposed method reduces the false positive rate from 6.41% to 2.56% (a 60% reduction) and increases precision from 93.59% to 97.44%. These results demonstrate that integrating structured semantic contours significantly enhances loop detection accuracy in outdoor SLAM.