Complex Human Activity Recognition Using Smartphone and Wrist-Worn Motion Sensors.

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Wearable Technology

Background:

• On-body motion sensor placement significantly impacts human activity recognition (HAR).
• Current HAR often relies on single sensor locations (e.g., phone in pocket), limiting recognition of hand-centric activities.
• Wrist-worn sensors are used for hand gestures, but often in isolation from other body-worn sensors.

Purpose of the Study:

• To evaluate the effectiveness of combining motion sensors at both wrist and pocket positions for HAR.
• To investigate the impact of sensor fusion and varying window sizes on recognizing diverse human activities.
• To identify optimal sensor configurations and window sizes for improved HAR, particularly for less repetitive tasks.

Main Methods:

• Utilized three motion sensors: accelerometer, gyroscope, and linear acceleration sensor.
• Collected data from sensors placed at both wrist and pocket locations.
• Evaluated thirteen distinct human activities using three different classifiers and seven window sizes (2-30 seconds).

Main Results:

• The combination of wrist and pocket sensor data significantly outperformed wrist-only data, especially with smaller segmentation windows.
• Increasing window size positively impacted the recognition of less repetitive activities (e.g., smoking, eating) unlike repetitive activities (e.g., walking).
• Proposed optimizations further enhanced activity recognition performance.

Conclusions:

• Sensor fusion from multiple body locations (wrist and pocket) provides richer contextual information for HAR.
• Adaptive window sizing is crucial for effectively recognizing both repetitive and less repetitive human activities.
• The study offers a publicly available dataset to promote reproducibility and further research in HAR.