A Novel Spatio-Temporal Network of Multi-channel CNN and GCN for Human Activity Recognition Based on BAN

To improve the accuracy of human activity recognition (HAR) based on body area network (BAN), a novel spatio-temporal network combining multi-channel convolutional neural network (CNN) with graph convolutional neural network (GCN) is proposed in this paper. Based on BAN including multi-sensors, our model takes advantage of graph topological structure to represent human activity pattern. The multi-channel CNN is firstly constructed to greatly enhance learning ability to capture local representative feature. This benefits to exploit the most representative temporal activity features from time-series sensor data, overcoming the limitation of a single CNN with poor learning ability. And then the GCN with excellect extraction ability to local spatial feature is employed to discover the most representative spatial activity feature associated with the extracted temporal activity features in non-Euclidean space. This could gather the most valuable spatio-temporal dynamic activity feature hidden in human activity for HAR with high-generalization. The open wearable action recognition database dataset including all 13 actiivty patterns is used to evaluate the effectiveness of our proposed model. The experimental results showed that, with the limited sample data, our model can achieve the best generalization (96.01% accuracy, 95.39% precision, and 96.51% recall) to accurately detect all 13 activity patterns when compared with the recent relevant models. This suggested that our technique could gain surperior spatio-temporal activity representation learning capability in Euclidean and non-Euclidean space based on the limited sample data than the compared models. Our study may provide an effective technical solution for HAR wuth high quality.