Deep Learning-Based Vehicular IMU Calibration

Developed a deep learning framework to detect vehicle stationary states, utilizing zero-velocity updates (ZUPT) and gradient descent optimization to calibrate accelerometer bias, scale, and cross-talk.

Problem

Vehicular dead reckoning using low-cost Micro-Electromechanical Systems (MEMS) Inertial Measurement Units (IMUs) suffers from rapid positional drift over time. This drift is primarily driven by accumulated sensor bias, scaling errors, and axis cross-talk under dynamic driving conditions.

Approach

• Zero-Velocity Detection: Designed and trained a Deep Neural Network (DNN) classifier using statistical features calculated over sliding windows of IMU data to predict motionless intervals during driving (e.g., traffic stops).
• Drift Mitigation (ZUPT): Applied Zero-Velocity Updates (ZUPT) whenever stationary states were detected, resetting accumulated velocity drift.
• On-the-Fly Calibration: Leveraged the detected stationary segments to execute a magnitude-based gradient descent optimization (using the ADAM optimizer) to dynamically refine accelerometer bias, scale, and cross-talk parameters relative to the local gravitational field.

Key Highlights

• Real-World Validation: Evaluated against high-accuracy GPS ground truth on real vehicular datasets.
• Accuracy Improvement: Significantly reduced positional root mean square error (RMSE) compared to raw dead reckoning.
• Embedded Compatibility: Structured to run efficiently, making it suitable for real-time state estimation on embedded vehicle hardware.