Lightweight 2D CNN Fusion for Human Activity Recognition Using Multi-Transceiver CSI Data on Edge Devices

Human activity recognition (HAR) using channel state information (CSI) has gained significant attention for its potential in various IoT applications. This work introduces a lightweight 2D convolutional neural network (CNN) fusion model that leverages CSI data collected by four ESP32-S3-DevKitC1 devices, configured as transceiver pairs based on the Wi-Fi IEEE 802.11n standard. The CSI data, capturing 166 sub-carriers across 300-450 packets per sample, is transformed into time-frequency heatmaps using discrete wavelet transform (DWT). The dataset comprises 1200 samples per sensing for each of the ten activities, including clapping, jumping, no people, punching, push-pull, rubbing hands, squatting, standing, twisting, and waving. The proposed 2D CNN fusion model utilizes data from both the front and side sensing systems, achieving an 84.5% test accuracy with a model size of 1.65 MB and an inference time of 72.4 milliseconds on a Raspberry Pi 5 edge module. Comparative evaluations with pre-trained CNN models, including DenseNet, EfficientNet, Inception, and ResNet, reveal that EfficientNetv2s attains the highest test accuracy of 69.28%, though with a much larger model size of 78 MB and a slower inference time of 245 milliseconds. The fusion approach proves highly efficient for real-time, edge deployable, and scalable, making it ideal for real-time IoT-based HAR applications.