The increasing presence of wearable devices (smartwatches, low-power sensors) has paved the way to massive collection of physiological data for health, wellness, and sports applications. However, analysis of these data must be accurate, fast, and energy-efficient: our group has developed deep learning frameworks and models specifically designed to work on mobile platforms, ensuring real-time activity classification—directly on the device without the need for external servers.
Motivations & Objectives
- Provide activity recognition solutions compatible with wearable platform computational constraints, without compromising accuracy and generalization.
- Enable intelligent and contextual analysis of human activities for telemedicine, sports, chronic prevention applications.
- Optimize the pipeline for local data processing, reducing reliance on cloud transmission (privacy, latency, energy saving).
Methods, Innovations & Results
-
Deep learning for human activity recognition: A resource efficient implementation on low-power devices
BSN Conference, IEEE 2016
Introduced a deep learning framework for activity recognition executed on-node, using signal spectrograms to achieve robust recognition against different orientations and hardware. Validated on real and laboratory datasets, the system achieves inference times and consumption compatible with continuous use. -
A Deep Learning Approach to on-Node Sensor Data Analytics for Mobile or Wearable Devices
IEEE Journal of Biomedical and Health Informatics 2017
Extends the method by integrating shallow and deep features, optimizing the pipeline for complex and varied activities (health, sports). Demonstrates local classification can surpass shallow or deep solutions alone. -
Real-Time Food Intake Classification and Energy Expenditure Estimation on a Mobile Device
Body Sensor Network Conference, 2015
Combined application for automatic food intake classification and energy expenditure estimation directly on wearables, showing framework robustness for dietary pattern analysis—important for preventive health.
Code & Application
- ActiveMiles, demos and prototypes: on-device code available upon request and collaboration. Portability and adaptability for various mobile platforms.
Team & Authors
- Daniele Ravì (PI, deep learning framework, data fusion, edge/wearable pipeline design)
- Charence Wong, Benny Lo, GZ Yang, B. Hussain (algorithm development, clinical tests, UK/Asia validation, code-data fusion)