Human Activity Recognition via Wi-Fi CSI and VAE-CNN-BiLSTM Hybrid Deep Learning

Human Activity Recognition (HAR) is vital across multiple applications, such as healthcare monitoring, smart home systems, and surveillance. Recently, Wi-Fi channel state information (CSI) has gained attention as a valuable data source for HAR, owing to its non-intrusive nature and capacity to capture detailed spatial information. This paper introduces a deep learning framework that combines variational autoencoders (VAEs), convolutional neural networks (CNNs), and bidirectional long short-term memory (BiLSTM) networks to achieve high-accuracy HAR using Wi-Fi CSI data. The proposed hybrid VAE-CNN-BiLSTM model utilizes spatial information encoded in CSI matrices to extract discriminative features crucial for recognizing different activities. The VAE component captures a compressed representation of the high-dimensional CSI data, enhancing the model's robustness and generalization. The CNN layers then capture spatial features from raw CSI data, identifying both local and global patterns, which are further processed by the BiLSTM layers to capture the temporal dependencies and dynamic patterns associated with human activities. To evaluate the model, we conducted experiments on a publicly available Wi-Fi CSI dataset for human activity recognition in both line-of-sight (LOS) and non-line-of-sight (NLOS) indoor environments, comparing its performance with other state-of-the-art deep learning models. The results indicate that the hybrid VAE-CNN-BiLSTM model achieves superior accuracy, with classification rates of 88.9% in LOS and 88.2% in NLOS scenarios, demonstrating its effectiveness in accurately distinguishing various human activities.

I. Introduction

In today's modern world, human activity recognition (HAR) has gained significant attention due to its potential applications in healthcare, smart homes, security systems, and human-computer interaction. Traditionally, HAR relied heavily on wearable devices or cameras [1], posing limitations in terms of privacy concerns. However, recent advancements in wireless communication systems, particularly Wi-Fi signals, have demonstrated the potential to recognize human activities accurately. Wi-Fi signals, which are abundantly present in nearly every indoor environment, provide valuable insights into the movements and behavior of individuals. These signals propagate through space and interact with the surrounding environment, including the human body. When individuals move or perform actions within the Wi-Fi coverage area, they introduce variations in the signal properties due to reflection, diffraction, and absorption phenomena. These variations can be detected by analyzing the received signal strength indicators (RSSIs) and Wi-Fi signal channel state information (CSI). The RSSI data is widely employed in active localization but is person-dependent and lacks precision in capturing signal changes during human movements, particularly when a person is not positioned directly between an Access Point (AP) and a Wi-Fi router. CSI provides detailed information by measuring amplitude and phase distortions for each antenna pair at every sub-carrier frequency. This enables distinct patterns in time domain variations to be leveraged for human activity recognition.