yugenjr/Speednet-V2

SpeedNet v2: Multi-Task Neural Network for Inertial Dead Reckoning

1. Model Name

SpeedNet v2 (speednet-v2)

2. Purpose

SpeedNet v2 is a lightweight multi-task neural network designed for vehicle forward-speed and yaw-rate estimation from noisy 6-DOF IMU sensor data. It operates during GNSS outage windows in Intelligent Dead Reckoning (IDR) navigation systems, providing real-time speed, yaw-rate, stationary state logits, and short-term velocity change estimates to bound inertial drift without cloud connectivity.

3. Architecture

• Input Representation: 4-second temporal window ($W = 40$ samples at 10 Hz sampling rate, 6 IMU channels: accel_x, accel_y, accel_z, gyro_x, gyro_y, gyro_z).
• Feature Extractor:
  • 1D Convolutional Layer: 32 channels, kernel size 3, padding 1, BatchNorm1d, ReLU.
  • 1D Convolutional Layer: 64 channels, kernel size 3, padding 1, BatchNorm1d, ReLU.
  • Bidirectional LSTM (BiLSTM): 1 layer, input size 64, hidden size 64 (bidirectional output dimensionality = 128).
• Shared Dense Bottleneck: Linear(128, 64) + ReLU.
• Multi-Task Heads:
  • Forward Speed Regression ($v_{\text{fwd}}$): Linear(64, 32) + ReLU + Linear(32, 1) + ReLU (enforces non-negativity).
  • Yaw Rate Regression ($\omega_{\text{yaw}}$): Linear(64, 32) + ReLU + Linear(32, 1).
  • Stationary Logit ($\text{logit}_{\text{stat}}$): Linear(64, 32) + ReLU + Linear(32, 1).
  • Auxiliary Velocity Change ($\Delta v$): Linear(64, 16) + ReLU + Linear(16, 1).

4. Outputs

1. v_fwd: Estimated vehicle forward speed ($m/s$).
2. w_yaw: Estimated vehicle yaw rate ($rad/s$).
3. logit_stat: Logit output for binary zero-speed / stationary classification.
4. delta_v: Auxiliary short-term longitudinal velocity change ($m/s$).

5. Deployment

• Formats Included:
  • ONNX (speednet_v2_w40.onnx & speednet_v2_w40.onnx.data): Opset 14, constant-folded, dynamic batch dimension.
  • TorchScript Lite (speednet_v2_w40.ptl): Traced for PyTorch Mobile / Native Android execution.
• Target Platform: On-device smartphone and embedded edge runtimes (ONNX Runtime Mobile / PyTorch Mobile).
• Cloud Dependency: None. Designed for 100% offline, real-time edge execution during GNSS loss.

6. Dataset

Trained and evaluated on the IO-VNBD (Inertial Outdoor Vehicle Navigation & Benchmark Dataset), featuring aligned high-frequency smartphone IMU telemetry paired with high-precision RTK GNSS ground truth across diverse driving trajectories.

7. Benchmarks

• Offline Trajectory Benchmark (M028 / M029 EKF System): $218.93\text{ m}$ position error after a 300-second complete GNSS outage.
• On-Device Edge Latency: Not yet benchmarked
• On-Device Energy Consumption: Not yet benchmarked
• Quantized Edge Throughput: Not yet benchmarked

8. Note on M028 / M029 Benchmark Configurations

The M028 and M029 benchmark configurations utilize the exact same SpeedNet v2 neural model weights (speednet_v2_w40.pth). The performance improvements observed in M029 over previous milestones stem entirely from enhancements to the navigation, jerk-gating, and heading filter integration architecture, rather than retraining or modifying the underlying neural network checkpoint.

9. Limitations

• Sensor Orientation & Calibration: Model predictions rely on consistent IMU axis mapping and pre-calibrated accelerometer/gyroscope bias offsets.
• Smartphone Domain Shift: Variations in vehicle mounting, vibration spectrums, and smartphone sensor noise characteristics across different hardware models may impact speed accuracy.
• Target Hardware Latency: Edge inference latency must be validated on the specific mobile processor target.
• Navigation Accuracy Scope: The published neural model artifacts alone do not guarantee full navigation accuracy; system performance depends on integration with an Extended Kalman Filter (EKF) or Dead Reckoning engine.