On the Design of Efficient Deep Learning Methods for Human Activity Recognition in Resource Constrained Devices