Lightweight Two-Stream Convolution-Augmented Transformer for Multi-Task Human Activity Recognition Using Wi-Fi Sensing

With the rapid growth of smart cities and smart homes, indoor human activity recognition (HAR) has become a critical technology for applications in security, health monitoring, and smart home automation. This paper introduces a lightweight and efficient HAR method based on Channel State Information (CSI) from Wi-Fi signals, utilizing the LTHAT network, a simplified yet powerful model optimized for resource-constrained environments. The proposed approach employs a modified Convolutional Neural Network (CNN) architecture with reduced computational complexity, featuring an adaptive attention mechanism and Gaussian encoding module for multi-label activity recognition. By leveraging the reduced attention heads and simplified encoder layers of LTHAT, we enhance the model’s performance in multi-user settings while maintaining accuracy and efficiency. Extensive experiments on the WiMANS dataset show that the LTHAT network outperforms conventional models, offering superior accuracy and real-time performance suitable for deployment on edge devices and IoT sensors. This work presents a scalable, practical solution for real-time HAR applications in smart environments.