Visual Memory Meets Geometric Filtering for GPS-Denied UAV Localization

Localizing an unmanned aerial vehicle (UAV) against a georeferenced satellite map can provide a position estimate when GPS is unavailable. A single aerial observation, however, may match multiple nearby map locations. We present \method, a reference-conditioned localization network that integrates evidence at three time scales. A gated recurrent unit (GRU) retains visual information across observations; a Kalman estimator propagates position and uncertainty between observations; and Mean Shift converts cross-view responses into continuous position estimates. The network first extracts a visual measurement with Mean Shift, fuses it with the recurrent features and predicted geometric state, and then applies a second Mean Shift step to refine the updated estimate. On 1,482 observations from eight held-out Bearing-UAV-90K routes, the full system achieves a mean localization error of 3.76 m and a 41.90% success rate within 3 m. Ablations show increased error when recurrent processing or the final refinement stage is removed. Evaluation assumes that a georeferenced map neighborhood is provided.