Hybrid Neural Network-Based PDR with Multi-Layer Heading Correction Across Smartphone Carrying Modes

Area of Science:

• Computer Science
• Robotics
• Navigation Systems

Background:

• Traditional pedestrian inertial navigation (PDR) algorithms fail to account for dynamic changes in smartphone carrying modes, impacting heading estimation accuracy.
• Existing carrying mode recognition methods lack robustness and universality, leading to significant localization errors during mode switching.
• Dynamic changes in carrying modes, such as during phone calls, introduce heading estimation errors crucial for accurate PDR.

Purpose of the Study:

• To develop an innovative PDR framework that overcomes limitations of traditional methods by addressing dynamic carrying mode changes.
• To improve heading estimation accuracy and reduce localization errors in PDR systems under practical, real-world conditions.
• To enhance the robustness and universality of carrying mode recognition for PDR applications.

Main Methods:

• Classified four common smartphone carrying modes and designed a CNN-LSTM hybrid model for real-time mode recognition with 99.68% accuracy.
• Implemented a multi-layer heading correction strategy, including a quaternion-based universal filter (VQF) for initial heading estimation.
• Developed algorithms for detecting mode switching points, adaptive offset correction for dynamic heading compensation, and heading optimization with lateral displacement constraints.

Main Results:

• The CNN-LSTM model achieved high accuracy in classifying smartphone carrying modes.
• The multi-layer heading correction strategy significantly reduced heading errors, with an average error below 1.5°.
• The proposed PDR framework demonstrated low cumulative positioning error (<1% of walking distance) and root mean square error (<2 m) in validation experiments.

Conclusions:

• The proposed PDR framework effectively addresses the challenge of dynamic carrying mode changes in smartphone-based navigation.
• The integrated approach of accurate mode recognition and multi-layer heading correction substantially improves localization accuracy.
• The framework shows significant potential for reliable pedestrian navigation in complex and dynamic real-world environments.