Smartphone-Based Activity Recognition in a Pedestrian Navigation Context.

Area of Science:

• Mobile computing
• Human-computer interaction
• Machine learning for activity recognition

Background:

• Smartphone navigation systems require user activity and device placement data for accurate positioning and context-aware assistance.
• Existing human activity recognition (HAR) methods often focus on general activities, lacking specificity for navigation contexts.
• Landmarks like stairs and elevators are crucial for indoor localization, necessitating context-specific recognition.

Purpose of the Study:

• To develop and evaluate machine learning models for recognizing specific user activities and device placements relevant to pedestrian navigation.
• To create a dataset encompassing diverse device placements and navigation-specific activities.
• To assess the performance and resource efficiency of hierarchical classifiers for enhancing navigation systems.

Main Methods:

• Collected a dataset of over 6 hours of sensor data across 28 combinations of device placements and activities.
• Utilized Long Short-Term Memory (LSTM)-based machine learning to train hierarchical classifiers.
• Evaluated classification accuracy for device placement and specific user activities.

Main Results:

• Achieved 97.2% accuracy in device placement classification, comparable to benchmarks but more resource-efficient.
• Activity recognition performance ranged from 62.6% to 98.7%, closely matching benchmark results.
• Demonstrated the practical application of hierarchical classifiers in a case study analyzing navigation behavior.

Conclusions:

• Hierarchical classifiers effectively distinguish between specific user activities and device placements for smartphone navigation.
• The developed methods offer a resource-efficient approach to enhance indoor positioning and context-aware navigation.
• The study provides valuable insights into real-world navigation behavior by correlating user activity and device placement.