A Comparative Analysis of Lightweight CNN Architectures for Wi-Fi CSI-Based Fall Detection

Fall detection for elderly care has gained substantial research attention, with Wi-Fi Channel State Information (CSI) emerging as a promising non-intrusive sensing modality that addresses privacy concerns associated with camera-based systems. However, the adoption of deep learning in this domain requires models that balance high accuracy with computational efficiency, enabling deployment on resource-constrained devices. This study presents a comparative analysis of four lightweight Convolutional Neural Network (CNN) architectures-MobileNetV2, MobileNetV3, EfficientNet-B0, and ShuffleNetV2-that has been presented for Wi-Fi CSI-based fall detection. The models were evaluated using classification metrics, including accuracy, loss, precision, recall, and F1-score, as well as efficiency metrics such as model size, parameter count, FLOPs, and inference time. Experimental results showed that the highest classification accuracy of 98.61% was achieved by EfficientNet-B0, along with the lowest loss and best F1-score (0.986), while a near-fastest inference time of 8.53 ms was maintained. This combination of predictive performance and computational efficiency has been demonstrated to highlight strong potential for practical, real-time fall detection applications. Meanwhile, ShuffleNetV2 achieved a comparable accuracy of 98.47% and an $F 1$-score of 0.985, with the fastest inference time of $\mathbf{7. 1 9}$ ms and a small model size of only 5.21 MB. These results indicate that ShuffleNetV2 provides an excellent trade-off between accuracy and computational cost, making it a highly suitable candidate for deployment on real-time edge devices.