PROJECT TITLE :

On the Personalization of Classification Models for Human Activity Recognition

ABSTRACT:

A considerable portion of recent literature on Machine Learning approaches has focused on automatic recognition of people's activities. The growing availability of equipment capable of acquiring signals that, when properly processed, can reveal information about human activities of daily living is the fundamental cause for this intense interest (ADL). Machine learning approaches that process information from wearable sensors and/or cameras strategically placed in the surroundings are commonly used to recognize human activities. Human activities have a strong subjective characteristic that is tied to various aspects such as age, gender, weight, height, physical ability, and lifestyle, regardless of the type of sensor used. Personalization models have been investigated to account for these subjective elements, and it has been established that the accuracy of machine learning algorithms may be enhanced by applying these models. In this paper, we look at how to recognize human actions using signals collected by a smartphone's accelerometer. This study primarily makes three contributions. The formulation of a clear validation model that takes into account the problem of personalization and thus allows for objective evaluation of machine learning algorithms is one of the earliest contributions. A second contribution is the testing of a personalization model on three separate public datasets that analyzes two aspects: physical similarity (age, weight, and height) and intrinsic features of the signals emitted by these persons when doing activities. The construction of a personalization model that analyzes both physical and signal similarities is the third and last contribution. The results reveal that using personalization models enhances accuracy on average, proving the validity of the method and paving the path for more research into this area.