|
|
 |
| |
|
|
|
Development of Intelligent 3D UAV Navigation Based on Inertial and Satellite Data |
|
|
|
PP: 1645-1658 |
|
|
doi:10.18576/amis/200612
|
|
|
|
Author(s) |
|
|
|
Ayan Kemel,
Sabyrzhan Atanov,
Luigi La Spada,
|
|
|
|
Abstract |
|
|
| Reliable three-dimensional navigation of unmanned aerial vehicles requires continuity when satellite positioning is temporarily unavailable. This paper presents Adaptive FusionNav, a causal recurrent model that combines inertial measurements with causally available Global Navigation Satellite System (GNSS) history and learns a bounded velocity attenuation gate for GNSS outages. Experiments use the real-flight INSANE dataset with flight-level separation into training, validation, and previously unused test sequences. Artificial GNSS outages of 10, 30, and 60 s are evaluated with three neural-network seeds. On held-out test flights, the IMU-enabled model achieved relative-motion RMSEs of 1.356, 8.972, and 17.222 m, compared with 2.055, 9.908, and 19.325 m for a parameter-matched GNSS-only neural model, corresponding to reductions of 34.0%, 9.4%, and 10.9%. The IMU-enabled variant was better in 17 of 18 test episode–seed comparisons. However, simple Held-GNSS and kinematic baselines remained stronger on the aggregate held-out test. The results therefore support a reproducible incremental contribution of inertial data within the matched learned fusion architecture, rather than universal superiority over classical baselines.
|
|
|
|
|
|
|
|
