WaveHAR: Learning Wavelet Representation for Wearable Human Activity Recognition

In the rapidly evolving Internet of Things (IoT) industry, Human Activity Recognition (HAR) technology based on wearable sensors always plays a pivotal role. Deep learning models, especially Convolutional Neural Networks (CNNs) have achieved remarkable success in the HAR domain. However, existing research still faces with two major limitations: 1) Small-kernel CNNs struggle to establish long-range dependency relationships when extracting motion features; 2) Traditional convolutional layers tend to focus on high-frequency information, neglecting the low-frequency component, which nonetheless contains rich feature information. Although some studies have addressed the limitations of small-kernel CNNs by adopting large-kernel CNNs, leading to significant performance improvements, the issues of over-parametrization and performance saturation associated with large-kernel CNNs cannot be ignored. To tackle these challenges, we introduce a novel method that expands the receptive field bycascading wavelet decompositions at multiple levels. This process enhances the model’s ability to capture low-frequency information while complementing traditional convolutional layers that primarily focus on high-frequency features. While wavelet transform is a well-established tool in signal processing, how to apply it to extend the receptive field in wearable sensor-based HAR has been rarely explored. Unlike previous wavelet research, this paper mainly targets the expansion of the receptive field in the low-frequency domain of small-kernel convolutional layers, aiming to enhance activity recognition performance. To thoroughly evaluate the proposed method, we conducted extensive experiments on multiple public datasets and compared it against several competitive state-of-the-art baselines. The experimental results consistently demonstrate that our method exhibits a clear advantage in a variety of activity recognition tasks, validating the effectiveness and efficiency of this approach.