ROCIP: robust continuous inertial position tracking for complex actions emerging from the interaction of human actors...

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Consumer-grade inertial measurement units (IMUs) are improving but suffer from noise and bias.
• Pedestrian dead reckoning (PDR) methods struggle with long-term error accumulation.
• Integrating inertial tracking with global reference frames remains a challenge.

Purpose of the Study:

• To develop an "error introspecting" system for inertial navigation.
• To overcome limitations of traditional PDR and machine learning approaches.
• To enable accurate, long-term self-contained inertial tracking.

Main Methods:

• A neural statistical motion model using DenseNet to predict poses and uncertainties.
• Integration with a Rao-Blackwellised particle filter (RBPF) for probabilistic calibration.
• Collection of a novel head-mounted IMU dataset with diverse motion patterns.

Main Results:

• The proposed ROCIP method outperformed leading approaches in inertial tracking.
• Achieved a relative trajectory error (RTE) of 4.94m and absolute trajectory error (ATE) of 4.36m.
• Demonstrated sustained accuracy and minimal error accumulation during long-term tracking.

Conclusions:

• ROCIP effectively addresses error accumulation in dead reckoning.
• The system offers a robust solution for long-term, self-contained inertial tracking.
• This advancement has broad applicability in areas requiring precise motion tracking.