This whitepaper evaluates convolutional neural network architectures trained on CSI amplitude spectrograms for device-free classification of human activities, comparing accuracy and latency against classical machine-learning baselines.
Describes the short-time Fourier transform pipeline converting raw CSI time series into Doppler spectrograms, and the labeled activity dataset used for supervised training across walking, sitting, falling, and idle classes.
Presents a compact CNN architecture achieving real-time inference on edge hardware, reporting classification accuracy, confusion matrices, and comparison against SVM and Random Forest baselines on held-out subjects.