AHARNet: Adaptive Human Activity Recognition Model for Multimodal Consumer Electronics With Different Computation Resources

With the development of consumer electronics, enhancing smart terminal devices to better serve humans has received increasing attention. Human Activity Recognition (HAR) plays a crucial role in enhancing and improving human life on smart terminal devices, and has become a popular research topic. HAR models based on deep learning have demonstrated impressive performance due to their powerful feature extraction capabilities. As HAR has rich application scenarios, HAR models typically need to be deployed on devices with different computation resources. However, the existing HAR models lack the adaptability to different computing resources. To address this issue, this paper proposes an adaptive HAR model with resource-aware capability, AHARNet. Our model has multiple output layers and achieves adaptive inference through an early-exiting mechanism. AHARNet first generates early-exiting decisions based on the available computation resources of the device through a computation resource perceptron. Then during inference, AHARNet only performs partial layers to meet the device’s resource constraints. To validate the effectiveness of our model, experiments are performed on three public datasets. The experimental results show that compared to the baseline architecture, AHARNet can dynamically reduce over 61% of computational overhead while demonstrating comparable accuracy.