Camera View Selection for Bearing-Only Geolocation of Stationary Targets Using Multi-UAV Swarms with GPS Bias

ABSTRACT: We address the problem of geolocating a static ground target from multiple low-cost Unmanned Aerial Vehicles (UAVs) equipped with monocular cameras and object detectors. Each drone generates large sequences of bearing-only measurements per flyover, but these are affected by per-drone GPS biases (systematic position errors specific to each UAV) that remain effectively constant over short trajectories. We propose a greedy selector that combines angular diversity and detection confidence to pre-filter measurements before Random Sample Consensus (RANSAC)-based robust fusion. Across 30 randomized simulation runs, the selector achieves mean accuracy comparable to full-pool RANSAC (2.34 m vs. 2.44 m) while reducing median processing time by 4.1× (393 ms vs. 1602 ms). Similar accuracy is observed against LO-RANSAC and PROSAC, with the selector remaining 4× to 35× faster. The selector also outperforms confidence-only and standard Fisher Information Matrix (FIM)-based subset baselines. A bias-aware FIM analysis via Schur-complement marginalization shows that the information contributed by repeated measurements from a single drone saturates under a shared position bias, explaining the single-drone concentration observed for these baselines. In the nominal setting, the angular-diversity proxy retains 99.7% of the bias-aware FIM objective and achieves localization error similar to direct FIM-based selection, while being 3.3× faster at the selection stage. Real flights with a low-cost quadrotor measure short-term GPS drift of 1.9–3.5 m over 10–63 min, supporting the assumed bias dynamics, and expose a failure mode in which persistent false positives capture the RANSAC consensus. Full-pool RANSAC shows slightly lower mean error at larger GPS-bias levels (σ_GPS ≥ 5 m), indicating that pre-selection is not preferable in every operating regime.