An investigation into non-invasive physical activity recognition using smartphones

Area of Science:

• Computer Science
• Biomedical Engineering
• Human-Computer Interaction

Background:

• Automatic monitoring of Activities of Daily Living (ADL) is crucial for identifying health deviations and evaluating interventions.
• Real-world activity recognition systems require unobtrusive sensing modalities that integrate seamlessly into human environments.
• Modern smartphones offer a ubiquitous and accessible platform for pervasive sensing applications.

Purpose of the Study:

• To investigate the feasibility of using smartphones as the primary sensing modality for activity recognition systems with limited placement constraints.
• To identify challenges associated with unconstrained sensor placement in smartphone-based activity recognition.
• To propose and evaluate solutions for improving the accuracy of activity recognition under unconstrained conditions.

Main Methods:

• Collected a dataset of 4 subjects performing 7 distinct activities under varying smartphone placement conditions.
• Employed a decision tree classifier for initial activity classification.
• Identified key problems related to unconstrained sensor placement and developed corresponding solutions.

Main Results:

• Initial experiments achieved precision and recall scores of 0.75 and 0.73, respectively, using a decision tree classifier.
• Identified 3 primary challenges associated with unconstrained smartphone placement for activity recognition.
• Implemented proposed solutions led to significant improvements: +13% in precision and +14.6% in recall.

Conclusions:

• Smartphones are a feasible and unobtrusive sensing modality for activity recognition in real-world health monitoring.
• Addressing challenges of unconstrained sensor placement is critical for robust and accurate activity recognition systems.
• The proposed solutions demonstrate a substantial enhancement in classification performance, paving the way for practical applications.