WiFiPi: Involuntary tracking of visitors at mass events

Abstract

Abstract-To simulate crowds at mass events, realistic movement data of people is required. Despite their limited capacity for approximating real human mobility, synthetic movement models are traditionally used for this purpose. More realistic simulations can be achieved by using real-life movement data, gathered by observing people in the desired context. This paper presents a method for tracking people at mass events without the need for active cooperation by the subjects. The mechanism works by scanning at multiple locations for packets sent out by the Wi-Fi interface on visitors' smartphones, and correlating the data captured at these different locations. The proposed method can be implemented at very low cost on Raspberry Pi computers. This implementation was trialed in two different contexts: a popular music festival and a university campus. The method allows for tracking thousands of people simultaneously, and achieves a higher coverage rate than similar methods for involuntary crowd tracking. Moreover, the coverage rate is expected to increase even further as more people will start using smartphones. The proposed method has many applications in different domains. It also entails privacy implications that must be considered when deploying a similar system.

Building up knowledge through passive WiFi probes

“…In such phase, each end device broadcasts a probe request management frame to stimulate in-range APs to manifest themselves (replying with a probe reply management frame). Such probe requests are usually broadcasted in sequence on all the available WiFi channels (1)(2)(3)(4)(5)(6)(7)(8)(9)(10)(11)(12)(13)(14). The set of information which is contained in (or can be easily extracted by) probe request frames include the Medium Access Control (MAC) address of the sending device, the Received Signal Strength Indicator (RSSI) out of the transmission of the frame and the Preferred Network List (PNL), that is a list of Service Set IDentifiers (SSID) of the WIFi networks which are already known by the sending device.…”

“…In [6], twelve WiFi sniffers are deployed in an area of about 340 m 2 . Fingerprint localization through k-Nearest Neighbour classifier is performed, with a reported median error of about 4.5 m. The work in [7] presents Probr, an open source software solution to capture and process probe requests in order to support several on-line analysis tasks, Trajectory estimation and crowd-control are also addressed in [10] and [12] with similar approaches. In [10], the authors propose an Hidden Markov Model which returns the most probable device trajectory receiving as input the WiFi detections of end devices.…”

“…Computer Communications 117 (2018) [1][2][3][4][5][6][7][8][9][10][11][12] average probe request emission frequency for idle terminals is in the range of one probe every 2-5 minutes. A similar behavior has been observed also in our datasets.…”

“…The model parameters are empirically estimated for the test environment through an exhaustive measurement campaign. Namely, for all the possible combinations of parameters p 0 and α, the position of the [1][2][3][4][5][6][7][8][9][10][11][12] reference user, that is the vector coordinates of the cell the user lives in, x , is then estimated by solving the following regularized least-square problem.…”

“…7 we set k equal to 2 in our case: the resulting cluster centers are the two feature vectors [1 0 0 0 0] and [1 1 0 1 0]. As one can see, the first cluster is [1][2][3][4][5][6][7][8][9][10][11][12] identified by users which are detected for a short period of time (the first bin is present), but never return. We denote this first cluster as the group of passers-by users.…”