Enhancing real-time heading estimation for pedestrian navigation via deep learning and smartphone embedded sensors

Area of Science:

• Computer Science
• Robotics
• Geomatics Engineering

Background:

• Smartphone-based pedestrian navigation accuracy is limited by sensor bias, thermal drift, and changing device orientation.
• Existing methods using pervasive or auxiliary resources have limitations in accuracy or seamless indoor-outdoor transitions.

Purpose of the Study:

• To enhance heading estimation accuracy for smartphone-based pedestrian navigation.
• To address challenges posed by sensor limitations and environmental interferences.

Main Methods:

• Leveraging low-voltage gyroscope (LVGO) measurements and self-recognized straight-line visual features from camera images.
• Developing a deep learning approach using a U-Net convolutional neural network for visual feature recognition and heading constraint.
• Fusing gyroscope and magnetic field data with visual heading constraints to mitigate drift and bias.

Main Results:

• The proposed method effectively mitigates accumulated gyro drift using LVGOs and visual features.
• Deep learning-based visual tracking of straight-line features enhances heading estimation precision.
• The method achieves a balance between recognition time delay and accuracy, enabling smooth real-time performance.

Conclusions:

• The enhanced heading estimation method improves accuracy and reliability in pedestrian navigation.
• This approach offers significant potential for assisting visually impaired individuals by leveraging features like tactile paving.
• Further testing with visually impaired users is recommended for application integration.