SmartFPS: Neural network based wireless-inertial fusion positioning system.

Area of Science:

• Robotics
• Computer Science
• Signal Processing

Background:

• Wireless-inertial fusion positioning systems commonly use empirical models and filters (e.g., Kalman, particle).
• These empirical models often lack accuracy in real-world scenarios, leading to increased positioning errors.
• Biases in predetermined parameters further exacerbate positioning inaccuracies.

Purpose of the Study:

• To propose a novel fusion positioning system utilizing an end-to-end neural network.
• To implement a transfer learning strategy to enhance neural network performance across varied data distributions.
• To address the limitations of empirical models in current wireless-inertial systems.

Main Methods:

• Developed an end-to-end neural network for wireless-inertial fusion positioning.
• Integrated a transfer learning strategy for improved model adaptability.
• Validated the system using Bluetooth-inertial positioning in a full-floor scenario.

Main Results:

• Achieved a mean positioning error of 0.506 m with the fusion network.
• Transfer learning improved pedestrian step length and rotation accuracy by 53.3%.
• Bluetooth positioning accuracy for diverse devices increased by 33.4%, reducing overall error by 31.6%.

Conclusions:

• The proposed neural network-based fusion system surpasses traditional filter-based methods.
• The transfer learning approach significantly boosts accuracy and adaptability in indoor positioning.
• This method offers a more robust solution for challenging indoor positioning environments.