Wi-Fi Device Identification in Crowd Counting Using Machine Learning Methods

Abstract

The increase in the availability of computational resources gave rise to new technologies to estimate the amount of people in a given area. In this context, algorithm-based solutions for crowd counting can be grouped into image-based and non-image based approaches, the latter considering any other feature that is not visual. Currently, due to the popularization of smartphones and mobile devices, several researchers have been using Wi-Fi request packets for crowd counting estimation. Assuming that, on average, each person in a given place carries a Wi-Fi device, the number of unique MAC addresses can be associated with the number of people. However, since the probe may capture all Wi-Fi traffic – which may include broadcast messages from other access points or packets from notebooks and desktop devices – some strategy must be applied in order to identify only personal mobile devices, thus improving the method accuracy. In this work, we trained classifiers to segment mobile from static devices through its Wi-Fi behavior pattern. Therefore, using data collected from different devices and in different environments, we evaluated the proposed methodology by using several machine learning algorithms. Best results were achieved with logistic regression and neural network (MLP). The results of this study suggest the feasibility of the proposed method for crowd counting in high-density Wi-Fi zones.