Activity graph based convolutional neural network for physical activity recognition using acceleration and gyroscope data

Recent deep learning technique has shown a strong ability of performing automatic feature learning and outperformed models fitting on hand-crafted features in HAR, but their performance heavily replies on large amount of labelling data. In this paper, we proposed a novel deep learning approach with optimal activity graph generation model for accurate HAR with multiple subjects using only acceleration and gyroscope data. In the approach, we designed a three-step sensor signal input sorting mechanism for generating an optimal activity graph containing alignments of neighbored signals in both width and height. Then, we proposed a deep convolutional neural network to automatic extraction of distinguishable features for HAR. The experimental evaluation was carried out on three public HAR datasets in comparing other state-of-the-art approaches. Our method averagely improved recognition accuracy about 5% compared with other methods on these three datasets, particularly suitable to accurate HAR with multiple subjects using limited sensing data.