Robust Human Activity Recognition via Wearable Sensors Using Dynamic Gaussian Kernel Learning

During the past decade, deep convolutional neural networks (CNNs) have shown remarkable success in ubiquitous human activity recognition (HAR) scenarios. However, prior most works are static, which have to manually predefine a fixed kernel size before training, hence requiring time-consuming human intervention. A more efficient solution is to determine the kernel size during training stage, rather than design stage. Different from previous static designs, this paper proposes an adaptive solution called Dynamic Gaussian Convolution (DgConv), which can adaptively learn an optimal kernel size over sensor data for each individual convolutional layer, without requiring laborious human intervention. We conduct extensive experiments on several popular benchmark HAR datasets, including PAMAP2, UCI-HAR, UNIMIB-SHAR, OPPORTUNITY, HAG, and WISDM. The results show that the proposed DgConv can consistently enhance model performance while maintaining similar inference-time costs, which shows a notable performance improvement compared to static convolution. In addition, our approach may improve the robustness of deep CNN classifiers to various types of noise sources. A practical implementation is evaluated on a real mobile device. In summary, our DgConv could benefit the neural architecture search for activity recognition from narrowing the search space via excluding kernel size, which will be explored in a future work.