Human Activity Recognition for Indoor Localization Using Smartphone Inertial Sensors

Area of Science:

• Computer Science
• Electrical Engineering
• Robotics

Background:

• Growing demand for location-based services and smartphone ubiquity drive interest in indoor localization.
• Human activities contain semantic information crucial for enhancing indoor positioning accuracy.
• Existing indoor localization systems can benefit from integrating activity recognition.

Purpose of the Study:

• To propose a deep-learning model for human activity recognition (HAR) using smartphone inertial sensor data.
• To integrate the developed HAR model into an existing indoor positioning system.
• To evaluate the impact of HAR on indoor localization accuracy, particularly for floor transitions.

Main Methods:

• Development of a deep-learning model utilizing a Convolutional Long Short-Term Memory (ConvLSTM) network.
• Training and validation of the HAR model on smartphone inertial sensor data to classify nine distinct activities.
• Integration of the HAR model's predictions into an indoor positioning system for real-world testing.

Main Results:

• The HAR model achieved accurate classification of nine human activities, including stationary, locomotion, and vertical movement.
• Integration of HAR into the indoor positioning system resulted in an average positioning error of 2.4 meters in a multi-story building.
• The inclusion of human activity information significantly reduced overall localization error and improved floor transition detection.

Conclusions:

• The proposed ConvLSTM-based HAR model effectively classifies human activities using smartphone sensor data.
• Incorporating human activity information into indoor localization systems enhances positioning accuracy and floor transition identification.
• This approach offers a promising method for improving the performance of indoor positioning technologies.