SASGeo: Semantic map boosts drone localization to 95% without GPS
New framework achieves 95% retrieval accuracy using roads and buildings as landmarks.
A new research paper from Natalia Trukhina and Vadim Vashkelis introduces SASGeo (Stability-Aware Semantic Map Localization), a framework designed to help drones navigate when GPS signals are unavailable. Instead of relying on raw visual appearance—which is brittle under changing seasons, lighting, or map age—SASGeo uses stable semantic features like roads, buildings, railways, and field boundaries. The system combines semantic raster alignment, relational graph evidence, and integrity-aware rejection of ambiguous fixes.
In 220 randomized synthetic trials involving rotation, scale changes, occlusion, and hard semantic decoys, a simple global semantic descriptor achieved only 58.6% Recall@1, while spatial semantic matching variants reached 94.5–95.5%. Wilson confidence intervals confirm the spatial approach is significantly better, though individual modules couldn't be statistically separated. The authors emphasize this is not yet validated in real flight, but the structured semantic geometry approach shows clear promise for GNSS-denied drone operations.
- SASGeo uses semantic features (roads, buildings, waterways) rather than raw appearance for UAV localization without GPS.
- Spatial semantic matching achieved 94.5–95.5% Recall@1 in 220 synthetic trials, vs. 58.6% for a global descriptor.
- The framework includes explicit handling of ambiguous fixes and integrity-aware rejection, but next steps require real-flight validation.
Why It Matters
Enables reliable drone navigation in GPS-denied environments, critical for search-and-rescue, mapping, and military operations.