Device-free Crowd Sensing in Dense WiFi MIMO Networks: Channel Features and Machine Learning Tools

The paper addresses the problem of passive crowd sensing in an indoor space by processing baseband radio signals originated from a dense WiFi network. Focusing on unmodified WiFi devices equipped with multi-antenna OFDM physical radio interfaces (IEEE 802.11n/ac), we investigate the selection of statistical features to measure the body-induced alterations of the channel state information (CSI) and we analyze their dependency over the space (antennas) and frequency domains. Different machine learning methods are compared and optimized to discriminate up to 5 people moving inside the smart space. We compare different solutions for classification and target counting based on feed-forward and recurrent neural networks based on long short term memory architecture (LSTM). Experiments with real subjects are conducted to validate the proposed approach. Results confirm that CSI feature selection is crucial to optimize the counting performance and space-frequency diversity needs to be exploited to provide high-accuracy sensing in complex indoor environments