Investigation of Environment Dependence in Wi-Fi CSI-Based Crowd Counting Systems

Crowd counting systems are generally categorized as vision-based and sensor-based, wherein the former requires LOS and may cause privacy concerns, and the latter requires contact and calibration. Recent systems explored the use of radio signals, particularly from Wi-Fi, to perform crowd counting and estimation. Wi-Fi crowd counting systems use channel state information (CSI) as input to the classifier to sense the number of people in the environment. These systems often do not investigate the environment dependence of the data (i.e., the training environment is the same as the testing environment). In this study, we gathered thousands of CSI time samples for 3 crowd count classes (0-person, 3-person, and 5-person) in various indoor environments for several days in the span of 3 months. Due to the high temporal and environmental dynamics caused by the 3-month long data collection, our crowd counting system trained without a prior knowledge of the target environment did not perform as well as the system trained with partial knowledge. By partially training the crowd counting model, we can see a performance boost of around 16–20% in the classification accuracy.