HT-AggNet: Hierarchical temporal aggregation network with near-zero-cost layer stacking for human activity recognition

Abstract

With the steady growth of sensor technology and wearable devices in pervasive computing applications, sensor-based human activity recognition has gained attention in fields such as healthcare monitoring and fitness tracking. This has resulted in an increased need for accurate and real-time systems. Recent studies to satisfy the real-time conditions have attempted to design lightweight neural networks by mainly restricting the number of layers shallowly, which has decreased both inference time and accuracy. To recover the loss of accuracy, we propose an innovative hierarchical temporal aggregation network (HT-AggNet) that allows the network architecture to be deeper, leading to an accuracy gain with only a near-zero increase in computational cost. Furthermore, a temporal glance convolution is presented to model the global context information of the signal patterns. Consequently, the HT-AggNet hierarchically extracts the local and global temporal information and then merges them based on hierarchical temporal aggregation. In our experiments, the HT-AggNet outperformed existing methods on seven publicly available datasets and achieved state-of-the-art performance. The source code for the HT-AggNet is publicly available at https://github.com/jgpark92/HT-AggNet.