RNN-Aided Human Velocity Estimation from a Single IMU.

Area of Science:

• Robotics
• Sensor Fusion
• Machine Learning

Background:

• Pedestrian Dead Reckoning (PDR) relies on inertial measurement units (IMUs) for positioning, but velocity estimation suffers from drift due to signal noise.
• Traditional PDR velocity estimation methods are application-specific, sensor-dependent, and require significant parameter tuning.

Purpose of the Study:

• To develop a robust and accurate velocity estimation method for PDR using a single, non-calibrated IMU.
• To overcome the limitations of classic PDR approaches and improve accuracy in dynamic movement scenarios.

Main Methods:

• A hybrid filter combining a convolutional neural network (CNN) for spatial feature extraction and a bidirectional recurrent neural network (BLSTM) for temporal relationship tracking.
• Integration of the CNN-BLSTM model with a linear Kalman filter (LKF) to refine velocity estimates.

Main Results:

• The proposed hybrid filter demonstrates robustness across various movement states and orientations, including highly dynamic situations.
• Outperforms conventional, machine learning, and deep learning methods, achieving velocity errors ≤0.16 m/s and distance errors ≤3 m/km.
• Exhibits excellent generalization to different and varying movement speeds, providing accurate and precise velocity estimations.

Conclusions:

• The novel CNN-BLSTM-LKF architecture significantly enhances PDR accuracy from a single IMU.
• This approach offers a more generalized and precise solution for velocity estimation in PDR compared to existing methods.