Real-Time Human Activity Classification Using Gait Cycle Averaging and Biometric Heuristics

The classification of human activities in real-time is an essential task of Human Activity Recognition (HAR). For the deployment of HAR systems on devices like smartphones and smartwatches, it is crucial to ensure their efficiency in terms of time and space. Within the realm of human activity classi...

Full description

Saved in:
Bibliographic Details
Published inProceedings (IEEE International Conference on Emerging Technologies and Factory Automation) pp. 355 - 361
Main Authors Ellison, Grant, Penelope Markovic, Milla, Yazdansepas, Delaram
Format Conference Proceeding
LanguageEnglish
Published IEEE 15.12.2023
Subjects
Activity Shapelets
Biometrics
Biometrics (access control)
Classification
Clustering
Human Activity Recognition
Machine Learning
Real-time systems
Sociology
Stairs
Systematics
Time Series
Time series analysis
Training
Online AccessGet full text

Cover

More Information
Summary:The classification of human activities in real-time is an essential task of Human Activity Recognition (HAR). For the deployment of HAR systems on devices like smartphones and smartwatches, it is crucial to ensure their efficiency in terms of time and space. Within the realm of human activity classification, a specific focus lies on the categorization of ambulatory activities, including walking, jogging, and ascending and descending stairs. Activity shapelets, which are geometric patterns representing the dominant pattern in ambulation, are extracted and used for accurate and efficient classification of an incoming time series signal from mobile devices. A trade-off must be made between an extensive training period for extracting shapelets tailored to each individual and the deployment of shapelets trained on a broader population at the expense of accuracy. We propose a novel approach for activity shapelet creation using gait cycle averaging, coupled with a method to partition subjects into training clusters based on biometric similarity. A systematic improvement in accuracy is shown when classifying activity data by leveraging biometric partitioning compared to randomly assigned training clusters. Our findings demonstrate that our methods can be used to deploy pre-trained shapelet libraries, eliminating the need for expensive individual training while maintaining high accuracy.
ISSN:1946-0759
DOI:10.1109/ICMLA58977.2023.00056