Eff-WHAR: A Lightweight Design for Efficient Wearable Sensor-Based Human Activity Recognition

Wearable sensor-based human activity recognition (WHAR) is an important application field that uses sensors for measurement, especially the broad application scenarios in human-computer interaction. Despite the strides made by deep learning-based approaches, practical issues such as deployment on edge devices are often overlooked due to their dependence on massive network layers that require extensive computational and storage resources. This article proposes a novel approach named Eff-WHAR, which aims to improve recognition accuracy and reduce computational burden in the context of WHAR, especially for practical applications on edge devices. The overall approach consists of two parts: multiscale temporal embedding network and channel mixing network. The first part involves utilizing convolution operations with varying receptive fields in the temporal dimension to extract and merge features from the input data. It facilitates a more comprehensive understanding of different behaviors, resulting in enhanced accuracy. In the second part, features are extracted and fused in the sensing channel dimension through a combination of intra-channel and inter-channel convolutions. It is better to reduce computational cost and model size by replacing the traditional 2-D convolution and attention mechanisms, which are commonly used but require a lot of computing power and storage, making them unsuitable for devices with limited resources. We validated Eff-WHAR on eight commonly used WHAR datasets, obtaining outstanding performance compared with SOTA approaches. In addition, we conducted experiments in actual edge devices, verifying the practicality and efficiency of our approach.