Adaptive Human Daily Activity Recognition Using Accelerometer Sensory Data from Smartphones

Abstract

Human activity recognition using streaming data from the accelerometer sensor of smartphone is still an interesting issue for researchers. Most researches develop the recognition model based on personal model type which require the training data obtained from only user who will utilize the model. To prepare the training data, the user must perform various activities and annotate them within the specified time. This is a major inconvenience for the users. In this paper, we propose a new smartphone-based dynamic framework for physical activity recognition named “ISAR+”. The new framework is an impersonal (universal) model which can be built once and used on new users without requiring labeled training data from those users. Because the proposed model is adaptability with evolving data streams of each new user by using the incremental learning for real-time recognition. This work was validated the proposed model in terms of prediction accuracy and usage times on real activity recognition datasets. In the experimental results, we show that ISAR+ can achieve the best performance for streaming activity recognition compared with the state-of-the-art models, especially across different users and without inquiry from users.