| --- | |
| license: mit | |
| tags: | |
| - onnx | |
| - time-series | |
| - imu | |
| - navigation | |
| - dead-reckoning | |
| # COAST-VNet-1 (Tiny) | |
| Edge-native, frequency-decoupled AI model for two-wheeler dead reckoning in GNSS-denied environments (urban canyons, tunnels, underpasses). | |
| ## Model Details | |
| - **Architecture:** Frequency-Decoupled CNN-GRU | |
| - **Parameters:** 96,086 | |
| - **Size:** 392 KB | |
| - **Format:** ONNX (Optimized for Mobile/Edge CPU) | |
| - **Input:** (20, 6) tensor — 2.0s window @ 10 Hz (ax, ay, az, gx, gy, gz) | |
| - **Output:** (6) — Forward speed, Yaw rate, Roll/Pitch residuals, and Log-Variance (Uncertainty). | |
| ## Training Data | |
| Trained on the IO-VNBD Dataset (23 drives). Ground truth labels were sourced exclusively from the vehicle CAN-bus, completely bypassing noisy phone GPS signals to ensure the model learns true kinematics. | |
| ## Use Case | |
| Designed to maintain navigation continuity for ride-hailing and emergency response fleets. The model mathematically separates low-frequency vehicle motion (lean/acceleration) from high-frequency road and engine vibration. | |
| ## Performance | |
| - **Inference Speed:** 8.3 microseconds per step (19,682 Hz on standard CPU). | |
| - **Accuracy Gain:** 2.02x lower median position error compared to naive dead reckoning over 43 real-world GPS outages. | |