WiFi-based human activity recognition through wall using deep learning

Wireless sensing is a promising method that integrates wireless mechanisms with strong sensing capabilities. The current focus of using WiFi Channel State Information (CSI) for human activity recognition (HAR) is the line-of-sight (LoS) path, which is mainly affected by human activities and is very sensitive to environmental changes. However, the signal on non-line-of-sight (nLoS) paths, particularly those passing through walls, is unpredictable due to the weak reflected signals destroyed by the wall. This work proposes a method to achieve high-accuracy wireless sensing based on CSI behavior recognition with low-cost resources by showing through-wall and wider-angle predictions using WiFi signals. The technique utilizes MIMO to exploit multipath propagation and increase the capability of signal transmission and receiving antennas. The signals captured by the multi-antenna are delivered into parallel channels with different spatial signatures. An RPi 4 B is attached to an ALFA AWUS 1900 adapter utilizing Nexmon firmware monitors and extracts CSI data with flexible C-based firmware for Broadcom/Cypress WiFi chips. Preprocessing techniques based on CSI are applied to improve the feature extraction from the amplitude data in an indoor environment. Furthermore, a deep learning algorithm based on RNN with an LSTM algorithm is used to classify the activity instances indoors, achieving up to 97.5% accuracy in classifying seven activities. The experiment shows CSI can achieve accurate wireless sensing in nLoS scenarios with extended antennas and a deep learning approach.