Daily Human Activities Recognition Using Heterogeneous Sensors from Smartphones

On-line understanding human activities can contribute solutions to some problems existing in the smart-city schema such as health-care, urban mobility, or security. Wearable sensors, especially sensors embedded in smartphones, turn to be good data streams for human activity recognition (HAR) tasks. Unfortunately, most of the existing methods are evaluated on small and fixed-size datasets, and lack of data sharing as well as classifiers re-training functions. These issues will lead to the challenge of unadaptable learning when facing problems of volume and variety of data. In order to tackle these problems, this paper proposes a new method with adaptive, interactive, and general-personal-model training components, and data sharing on the cloud. The major advantage of the proposed method is very fast to detect human activities of a new user at the beginning (i.e. deploying a system to a new user) with an acceptable accuracy of detection using the general model. Then, the personal model will help to increase the accuracy of activities detection personally by interacting with users. Another advantage of the proposed method is to share data (e.g. sensory data, models, activities, and user's profiles) among users/apps joined the system. These data will help to increase the accuracy of models timely by re-training periodically. Besides, the method can be used as a human-activity sensor that streams detected human activities to related components of smart-city scheme. The proposed method is evaluated and compared to de-facto datasets as well as state-of-the-art of HAR using smartphones. The experimental results show that there is a significant improvement of HAR's accuracy when utilizing the proposed method.