Human Activity Recognition Based on Symbolic Representation Algorithms for Inertial Sensors

Area of Science:

• Computer Science
• Human-Computer Interaction
• Machine Learning

Background:

• Mobile sensing enables human activity recognition (HAR) on smartphones to understand behavior.
• Current HAR solutions face limitations due to smartphone computational resource constraints.
• Existing shallow and deep learning strategies often have high computational costs, hindering smartphone implementation.

Purpose of the Study:

• To evaluate alternative low-computational cost strategies for HAR on smartphones.
• To develop HAR solutions with reduced memory and processing requirements.
• To assess the feasibility of symbolic representation algorithms for mobile HAR.

Main Methods:

• Utilized Symbolic Aggregate Approximation (SAX) and Symbolic Fourier Approximation (SFA) for data representation.
• Evaluated classification algorithms suited for symbolic data: SAX-VSM, BOSS, BOSS-VS, and WEASEL.
• Conducted experiments on benchmark databases: UCI-HAR, SHOAIB, and WISDM.

Main Results:

• Symbolic representation algorithms demonstrated faster feature extraction (average 84.81% speedup).
• Significant reduction in memory consumption (average 94.48% decrease).
• Achieved accuracy rates comparable to conventional HAR algorithms.

Conclusions:

• SAX and SFA provide efficient and effective alternatives for mobile HAR.
• These symbolic methods overcome computational limitations of current smartphone-based HAR solutions.
• The evaluated symbolic approaches offer a promising direction for resource-constrained HAR applications.