Alternative Deep Learning Architectures for Feature-Level Fusion in Human Activity Recognition

In this paper, we propose new deep learning architectures to fuse data provided by multiple sensors. More specifically, we combine classical features extracted from a sensor and raw data of other sensors. In order to make this data fusion possible, we exploited convolution, dense, and concatenation layers. The Mobile HEALTH dataset has been used to support our study. The results show that the proposed architectures are suitable for future use in the Human Activity Recognition (HAR) domain since their performances are comparable or better than those presented in the recent literature and the reference architectures. Indeed, we reached approximately an accuracy of 0.965 and 0.995 for the leave-one-subject-out and train-test strategies, respectively.