Fusion of Signal-Based Multichannel CNN-LSTM Models for Human Activity Recognition using Wearable Sensors
Darshil Bagadia, Thitiphoom Teptit, Punthita Dabthong, Seksan Laitrakun · 2021
Human activity recognition (HAR) is one of the emerging fields under machine learning that has received multiple technological advancements recently. HAR monitors and recognizes human motions or activities and classifies them. Multichannel deep learning (DL) models have proven to extract outstanding levels of performance out of HAR. In this paper, we investigate fusion approaches to combine the outputs from multiple channels whose inputs are the signals from wearable sensors attached to different parts of the body. Each channel is based on a hybrid of one-dimensional convolutional neural networks and long short-term memory. According to where the fusion takes place and which channel outputs are fused together, we study and investigate six fusion approaches. Their classification performances are evaluated and compared by using well-known public HAR datasets. The best fusion approach achieves 98.77% and 96.51% in terms of accuracy when using the PAMAP2 dataset and the DaLiAc dataset, respectively.