Motion-State-Aware Adaptive Step-Length Smartphone PDR for GPS-Denied Pedestrian Localization

Area of Science:

• Computer Science
• Robotics
• Geomatics Engineering

Background:

• Smartphone-based pedestrian dead reckoning (PDR) offers infrastructure-free 2D localization in GPS-denied environments.
• PDR's open-loop nature leads to accumulated step-length and heading errors, exacerbated by changing pedestrian actions and phone carrying modes.
• Conventional PDR methods often use fixed step-length models, failing to adapt to dynamic user behaviors.

Purpose of the Study:

• To propose a novel motion-state-aware PDR method that enhances localization accuracy by adapting to user actions and carrying modes.
• To develop adaptive models for the Weinberg coefficient to improve step-length estimation.
• To enhance heading estimation continuity under magnetic disturbances.

Main Methods:

• A random-forest classifier identifies joint motion states (action type and carrying mode) from smartphone sensor data.
• Two adaptive variants for the Weinberg coefficient are proposed: a state-wise linear model and a Transformer-enhanced extension.
• Heading is estimated by fusing gyroscope increments and magnetometer observations.

Main Results:

• The proposed motion-state-aware PDR method demonstrates improved accuracy compared to fixed-parameter and established adaptive baselines.
• Diagnostic metrics and trajectory-level evaluations confirm the effectiveness of the adaptive coefficient modeling.
• The method shows robustness in environments with frequent motion-state transitions and magnetic disturbances.

Conclusions:

• The motion-state-aware PDR approach effectively mitigates accumulated errors by adapting step-length estimation to diverse user behaviors.
• Adaptive modeling of the Weinberg coefficient is crucial for accurate PDR in dynamic environments.
• Future work may explore applications in more complex scenarios like underground or multi-floor environments.