CSI-Based Human Activity Recognition via Lightweight CNN Model and Data Augmentation

Human activity recognition (HAR) involves detecting users' actions through the analysis of sensor and other data types using AI. While cameras and sensor data are often limited by privacy concerns and technical limitations, researchers have explored device-free solutions. These solutions leverage Wi-Fi signals, which are notably influenced by human movements at the level of channel state information (CSI). This article proposes HAR-LightCNN, a CSI-based HAR solution. The main component of this solution is a deep lightweight convolutional neural network (CNN), notable for its reduced computational demands. The network achieves a balance between depth and trainable parameters, aiming at facilitating real-time activity recognition without compromising performance. We enhance the model's generalization capabilities using time series data augmentation (DA) techniques, which help address the small-sized dataset and class-imbalance problems. Upon evaluation with unseen testing data, our method demonstrates high accuracy in single-user activity recognition, surpassing existing state-of-the-art approaches.