Realistic-Motion Activity Recognition
- Author(s): Mortazavi, Jack Bobak
- Advisor(s): Sarrafzadeh, Majid
- et al.
Sedentary behavior starting from a younger and younger age has led to many serious medical conditions including obesity, cardiovascular diseases and diabetes. These conditions have led not only to a significant health burden but a significant economic burden in treating such ailments. Advances in sensor technology have let us not only monitor human activity but use such activity in applications to help address this sedentary life style. Exergaming is the convergence of using physical activity monitoring techniques and video game design in order to create healthy video game activities.
This dissertation investigates the design and implementation of such exergames, culminating in the creation of one such game that encompasses multiple pieces to address the obesity epidemic and the several parallel conditions that must be met. New classification algorithms must be created to identify these fine-grain movements, often in a large, multi-class setting. The analysis of such an algorithm must incorporate two considerations, the accuracy as well as the responsiveness of such a system. This work investigates developing such an algorithm for these fine-grain detailed motions, achieving new, high levels of accuracy, as well as the appropriate trade-off between the complexity and accuracy of such systems and the responsiveness to detecting the appropriate movements in a responsive fashion, leveraging contextual information to strengthen the classification results for a user-centric recognition experience. Further, clinical trials are run to investigate the energy expenditure levels of such a system, showing and guaranteeing levels of energy expenditure that can promote a more active and healthy gaming experience that is still enjoyable to the user.
The methods investigated in this dissertation show an activity level of moderate activity, and regression techniques that can predict these activity levels within an error of at most only $1$ metabolic equivalent of task. User-specific models can improve the computational complexity and accuracy of such systems, reducing the delay needed to classify and improving classification results by about $12\%$ when training a model for a specific user. Further, the multiple model approach investigated helps improve classification of difficult, fine-grain activities with greater than $90\%$ accuracy and F-score, creating an end-to-end system for detailed physical motion recognition and a complex wireless health system and application.