Terrain-Prior-Guided Change Detection and Incremental Update for Long-Term LiDAR Map Maintenance

Accurate and robust change detection is critical for long-term autonomous navigation in dynamic environments. In this paper, we propose a terrain-prior-guided framework for efficient LiDAR map change detection. Our method first partitions the global map into overlapping submaps along the vehicle trajectory, enabling localized and redundant observation comparison. Each submap encodes Bird’s Eye View (BEV)-grid-based statistical features, which are used to perform residual comparisons with incoming observations. To reduce false positives from occlusion and sensor noise, a simplified raycasting mechanism is employed to construct occupancy grids and determine change candidates. Furthermore, a terrain prior is constructed from historical elevation information to filter out changes occurring in untraversable or irrelevant areas. Experimental results on a real-world autonomous driving platform demonstrate that our method accurately detects additions, deletions, and environmental updates while significantly reducing communication overhead. Compared with baseline approaches, our full pipeline reduces false positives by over 80% and decreases transmitted data by more than 80%, showing strong potential for long-term, low-bandwidth collaborative mapping.