AbstractThe lack of physical activity is often related to cardiovascular diseases. Simple exercise like walking or climbing the stairs can already decrease the risk of this type of diseases considerably. Activity recognition can play an important role in this topic. Finding an easy way to exactly monitor our daily behavior gives us a complete insight in our daily routines and can motivate us to exercise. Modern smart phones are equipped with a rich set of sensors and are a new alternative platform for activity recognition. But smart phones are worn at different locations and orientations. People may carry their smart phone in the front pocket of their trousers or in their purse or bag. In this thesis we try to recognize basic activities like walking, biking or running and estimate walking speed. We used two different smart phone locations (front pocket trouser and purse), without telling the subject how to store the smart phone. One solution to cope with the above mentioned viabilities is to use orientation independent features and location specific models. We show that activity can be recognized with a feature set that only depends on a person's motion variation, with only a minor decrease in performance. We also investigated the accuracy of three different speed estimation algorithms under these unconstrained conditions. In this study we show that the main performance accuracy depends on step frequency performance. An improved step frequency algorithm is proposed which improves all walking speed estimation algorithms.
|Date of Award||31 Dec 2013|
|Supervisor||O.D. Amft (Supervisor 1) & M. Altini (Supervisor 2)|
Habitual behavior monitoring using smartphones: impact of sensor input and personalized models on walking patterns
van Dort, M. P. (Author). 31 Dec 2013
Student thesis: Master