AUTO-HAR: An adaptive human activity recognition framework using an automated CNN architecture design

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in time-series analysis for Human Activity Recognition (HAR).
• Manual design of CNN architectures for HAR is complex, time-consuming, and prone to errors.
• Neural Architecture Search (NAS) offers automated optimization, overcoming human limitations.

Purpose of the Study:

• To develop an efficient automated framework (AUTO-HAR) for designing optimal CNN architectures for HAR tasks.
• To leverage evolutionary algorithms, specifically the Genetic Optimization Algorithm (GA), for NAS in HAR.
• To introduce a novel encoding schema and an expanded search space for improved architecture discovery.

Main Methods:

• Utilized the Genetic Optimization Algorithm (GA) for automated CNN architecture selection.
• Proposed a novel encoding schema and an expanded search space for HAR-specific NAS.
• Evaluated the AUTO-HAR framework on three benchmark datasets: UCI-HAR, Opportunity, and DAPHNET.

Main Results:

• The AUTO-HAR framework achieved high accuracy in human activity recognition.
• Average accuracies reached 98.5% (±1.1%) on UCI-HAR, 98.3% on Opportunity, and 99.14% (±0.8%) on DAPHNET.
• The proposed method demonstrates efficient and effective HAR through optimized CNN architectures.

Conclusions:

• AUTO-HAR successfully automates the design of high-performance CNNs for HAR.
• The GA-driven NAS approach with a novel search space enhances HAR accuracy.
• This framework offers a robust solution for complex HAR challenges, surpassing manual design limitations.