Indoor people density sensing using Wi-Fi and channel state information Mohan Liyanage 1* , Chii Chang 1 , Satish Srirama 1 , Seng Loke 2 1 Mobile & Cloud Lab, Institute of Computer Science, University of Tartu, Tartu 50090, Estonia 2 School of Information Technology, Deakin University, 221 Burwood Highway, Burwood VIC 3125, Australia Corresponding Author Email: liyanage@ut.ee https://doi.org/10.18280/ama_b.610107 Received: 15 March 2018 Accepted: 29 March 2018 ABSTRACT Device-free passive crowd estimation technologies are capable of measuring the density of people in an area, using existing wireless network infrastructure. It has been applied in various application domains such as pedestrian control, crowd management in subways, guided tours and so forth. In this work, we have designed, implemented and validated a device-free indoor human crowd density sensing method based on Channel State Information (CSI) captured by a single Wi-Fi receiver. We investigate the behaviour of the CSI amplitude variance of each receiving stream over the different subcarriers and propose a method to aggregate the CSI amplitude over time without losing critical information. Further, we evaluated the method using three different machine learning algorithms. The result shows the proposed method achieves an estimated accuracy of 99.8% with the Weighted K-Nearest Neighbour. Keywords: channel state information, crowd estimation, device-free, RF sensing, Wi-Fi 1. INTRODUCTION The process of estimating the number of people in a given area has become a significant research area over the past years. Robust crowd counting in either an open or closed environment is an important and also a challenging task. A lot of research has been done in the development of crowd density estimation systems and they have been applied in a wide range of applications, such as counting people in a festival [1], pedestrian control, crowd management in subways [2], and customer count estimation in retail shops and so forth. The most common and traditional approaches are vision- based systems that analyse a video or images to estimate the number of people in the scene [2-5]. Mostly, those image processing techniques are able to estimate the population effectively. However, these methods suffer from some inherent drawbacks. For example, environmental factors such as light, fog, and dust greatly affect the quality of the video/image. Further, installing cameras in public areas raises privacy concerns and also requires additional costs. Emerging smartphones have been utilised in many sensing scenarios such as interacting with proximal Internet of Things (IoT) to assist ubiquitous computing applications [6], utilising the inbuilt sensing mechanism to provide environmental sensor data to remote clients [7-8] or brokering the environmental sensor data from proximal sensors to remote data centres on the move [9]. Furthermore, establishing an ad- hoc network among smartphones is not limited to performing data routing or distributed processing [10], with the inbuilt proximal networking mechanisms (e.g. Bluetooth) or audio processing mechanisms, the smartphones can sense the density of people in proximity [1, 11-12]. Although the smartphone-based sensing approaches are promising, they rely on the participants to carry the specific devices. Different from the device-specific approaches, device-free Radio Signal Strength (RSS)-based human density sensing approaches, which use RSS as an indication of the signal propagation strength, do not require people to carry devices. The related frameworks have been proposed in various applications such as localisation [13], motion detection [14], human activity recognition [15], and crowd estimation [16-17]. Moreover, researchers also estimate the crowd density based on analysing RSS values of RFID tags [18-19] or ZigBee wireless nodes [20]. However, most existing RSS measurement-based models face a critical challenge in accuracy because of the fundamental problems of RF wave propagation in the indoor environment. There are possibly multiple signals arriving at the receiver through multiple paths, and also attenuated by reflection when the signal hits the surface of an obstacle. Consequently, the time- varying nature and, one Receiving Signal Strength Indication (RSSI) value per packet cannot establish an accurate prediction model in complex environments. In contrast to having only one RSSI value per packet, current widely used Orthogonal Frequency Division Multiplexing (OFDM) systems explore the fine-grained physical layer information in multipath environments. Different from RSS, Channel State Information (CSI) is a complex matrix of values from the physical layer where data are modulated on multiple subcarriers at different frequencies and simultaneously transmitted over the IEEE 802.11n Wi-Fi link [21]. The CSI consists of amplitude and phase shift information that describes how the propagated signal experienced by the different effects of scattering, fading, and power decay for each spatial stream on every subcarrier. According to this characteristic, CSI is more stable and has more information than RSSI and also we can get more information that is sensitive to the environmental variance. Recently, due to the robust nature of the CSI in complex environments, there are many passive-sensing (device free sensing) frameworks that have been proposed in several Advances in Modelling and Analysis B Vol. 61, No. 1, March, 2018, pp. 37-47 Journal homepage: http://iieta.org/Journals/AMA/AMA_B 37