Long-Term Distant Landmark Tracking for UAV Navigation without Satellite Positioning

Abstract

Navigation of unmanned aerial vehicles (UAVs) in Global Navigation Satellite System (GNSS)-denied environments remains a challenging problem due to the drift accumulation of inertial sensors and the limitations of conventional visual navigation techniques. This paper presents a navigation correction approach based on the detection and persistent tracking of distant visual landmarks that remain observable over extended portions of a flight trajectory. The proposed framework combines synthetic dataset generation, object detection, monocular depth estimation, and multi-object tracking to estimate heading deviations from visual observations. A synthetic dataset containing 2,000 images of distant natural and artificial landmarks was generated under diverse weather, illumination, and observation conditions. The navigation pipeline employs a YOLO-based detector, the Depth Anything depth estimation model, and the Deep SORT tracking algorithm to identify landmarks, estimate their relative distance, and maintain stable observations over time. Preliminary experiments demonstrate that distant landmarks can provide reliable visual references for heading stabilization during long-range UAV flights. The obtained results indicate the feasibility of using persistent visual landmarks as a complementary navigation source in environments where satellite positioning is unavailable or unreliable. The proposed approach provides a foundation for future research on long-range visual navigation and multi-sensor autonomous UAV systems.