Elevated Road Network (ERNet): A Lightweight Industrial‑Grade Neural Network for Elevated Road and Yaw Detection in Mobile Navigation

Mobile navigation requires accurate yaw detection, especially in elevated road areas where GNSS accuracy is low.

Highways and parallel service roads cause difficulty distinguishing whether a vehicle is on an elevated road.

Gaode Map Online Engine Center proposes Elevated Road Network (ERNet), a lightweight industrial‑grade neural network that uses four high‑semantic features: Satellite Plane Projection (SPP), group ID, elevated‑road distance, and sequential speed.

During inference, ERNet predicts a 10‑dimensional embedding C for a vehicle position and compares its distance to two learned road descriptors A (on‑bridge) and B (off‑bridge); the smaller distance determines the road state. A confidence‑constraint further refines the decision.

Training uses a massive automatically labeled dataset (≈1 billion samples) with triplet loss and batch‑all sampling within each group. Group embedding encodes local context; SPP features are computed by projecting satellite signals onto a hexagonal grid (19 cells × 4 satellite types × 3 statistics = 228 dimensions).

ERNet’s architecture consists of 10 weight layers with shared parameters for SPP across hexagons, and a 50‑dimensional group embedding.

Experiments on Beijing, Shanghai, and Guangzhou show ERNet achieves the highest accuracy for elevated‑road recognition (precision≈99.1 %, recall≈99.7 %) and yaw detection (accuracy > 99 %) compared with baselines such as map‑matching, GNSS‑rule, XGBoost, and variants of ERNet.

Future work includes extending ERNet nationwide, incorporating road topology, and reducing the detection latency (currently 15 s) by exploring recurrent neural networks.