Long-Term Activity Recognition
Capturing what you're doing, and when, for weeks
Activities play a key role in the way we structure our lives. The type of activity, how it is performed, and with whom, can reveal a person's intention, habit, fitness, state of mind, and social connectedness; It is therefore not surprising that a range of research fields, from human-computer interaction to medical- and cognitive sciences, display growing interest in automatic activity recognition.
The purpose of this project is to formulate algorithms that can be embedded in small wearable devices to process sensor data in an energy- and memory-efficient way, while capturing the essence of the human activity. The targeted approach is to convert the sensor data immediately to detected activities onboard the worn sensor, rather than storing all raw sensor values, which would lead to high energy and storage requirements on the wearable device. Apart from algorithms, this project will produce dedicated hardware and closely annotated benchmark data, and involves researchers with vested interest in practical activity detection applications.
Present-day sensors and microcontrollers already permit power efficient operation and wearable-sized sensors, mainly due to a growing prevalence of personal computing products. Novel algorithms to learn and detect human activities from on-body sensor data have been experimented on as well, though mostly for short spans of time.
This project aims at advancing this research by working toward recognition of human activities over sustained periods of time, using tiny wearable sensors that detect motion- and posture characteristics, and map these to human physical activities.
An example of the data captured by the HedgeHog sensor when wrist-worn: Fullscreen Example.