Enhancing Dynamic Human Activity Recognition Through a Novel Martingale-Based Algorithm for Change Detection

The escalating prevalence of diseases linked to physical inactivity, including cardiovascular diseases, hypertension, obesity, diabetes, colon cancer, anxiety, depression, lipid disorders, and osteoporosis, poses a formidable challenge to modern healthcare. Despite advancements in wearable sensor technologies facilitating the study and monitoring of physical activities for improved well-being, existing methods for human activity recognition grapple with noise-related issues, impacting result accuracy. In this paper, we introduce a groundbreaking approach to Human Activity Recognition (HAR) by integrating martingale methods with smoothing, heuristic thresholding, and optimisation techniques. Our method addresses the pressing challenge of accurately identifying and estimating points of interest, such as Physical Activity Bout (PAB) duration, in HAR sequences. The unique contribution lies in our method’s ability to capture intricate patterns and dependencies within these sequences, leading to significantly improved accuracy compared to traditional approaches. With an impressive 93.40% accuracy and 90.4% G-mean measure, our method surpasses existing methods like multivariate randomised power martingale, multivariate exponential weighted moving average, mean absolute deviation, and extreme learning machine methods. Moreover, this research underscores the urgency of addressing physical inactivity-related health challenges and offers a pioneering solution with substantial performance enhancements. Additionally, our novel martingale-based approaches have practical implications for real-time monitoring and interventions aimed at promoting physical activity and mitigating associated health risks.