LTE-ACN: A Human Activity Recognition System Using LTE Signals and Attention-Guided ConvNeXt Networks

WiFi Channel State Information (CSI)-driven Human Activity Recognition (HAR) is widely researched for applications in activity monitoring and healthcare. Despite the proposal of numerous methods, persistent challenges such as high deployment costs, system instability, and limited sensing range continue to hinder progress. To address these issues, in this paper, we propose a novel technique called LTE-ACN (LTE Attention-guided ConvNeXt), which leverages commercial Long-Term Evolution (LTE) technology for human activity recognition. Firstly, we modify the open-source software srsRAN to obtain Cell Reference Signals (CRS) and collect real-time CSI from LTE base stations. This allows us to extract intrinsic features from the CSI observations. Next, due to the long distance between LTE base stations and terminals, signal propagation encounters various obstacles, leading to significant interference. To tackle this, we employ refined signal processing techniques, including outlier removal, noise filtering, Savitzky-Golay smoothing, peak-valley enhancement, and Gramian Angular Field (GAF) transformation. These GAFs are then subjected to Contrast Limited Adaptive Histogram Equalization (CLAHE) for image enhancement, resulting in the construction of a comprehensive human activity dataset. The enhanced GAF images are fed into an integrated ACN (Attention-guided ConvNeXt), which utilizes ConvNeXt to minimize information loss and extract more feature information. Additionally, the network incorporates the Convolutional Block Attention Module (CBAM) attention mechanism to enhance the relationship between each feature within the spatial domain. Experimental results demonstrate that our proposed technique significantly improves recognition accuracy and robustness, achieving an average recognition accuracy of 96.44% for human activities, while exhibiting superior adaptability and computational efficiency.