Activity classification using realistic data from wearable sensors

Area of Science:

• Human-computer interaction
• Biomedical engineering
• Machine learning for health

Background:

• Automatic classification of daily activities aids in promoting health-enhancing physical activities and healthier lifestyles.
• Wearable sensor technology offers a promising avenue for objective physical activity monitoring.

Purpose of the Study:

• To investigate methods for recognizing everyday activities such as walking, running, and cycling.
• To determine the utility of various sensors and the necessary signal processing and classification techniques.
• To establish a comprehensive dataset for activity recognition research.

Main Methods:

• Collected a large dataset of sensor data from 16 participants performing everyday activities.
• Utilized a 35-channel data recording system with wearable sensors over approximately 31 hours.
• Employed three distinct classification algorithms: custom decision tree, automatically generated decision tree, and artificial neural network.
• Performed leave-one-subject-out cross-validation to assess classifier performance.

Main Results:

• Classification accuracies varied by classifier and activity, ranging from 58% to 97% for decision trees and 22% to 96% for artificial neural networks.
• The automatically generated decision tree achieved the highest total classification accuracy at 86%.
• Custom decision tree and artificial neural network classifiers showed total accuracies of 82%.

Conclusions:

• Automatic activity recognition using wearable sensors is feasible and effective for promoting physical activity.
• Decision tree and artificial neural network models demonstrate strong potential for classifying everyday activities.
• Further research can refine sensor selection and classification algorithms for improved accuracy and broader application.