Counting targets (people or things) within a monitored area is an important task in emerging wireless applications, including those for smart environments, safety, and security. Conventional device-free radio-based systems for counting targets rely on localization and data association (i.e., individual-centric information) to infer the number of targets present in an area (i.e., crowd-centric information). However, many applications (e.g., affluence analytics) require only crowd-centric rather than individual-centric information. Moreover, individual-centric approaches may be inadequate due to the complexity of data association. This paper proposes a new technique for crowdcentric counting of device-free targets based on unsupervised learning, where the number of targets is inferred directly from a low-dimensional representation of the received waveforms. The proposed technique is validated via experimentation using an ultra-wideband sensor radar in an indoor environment. Index terms-Crowd-centric counting, unsupervised learning, dimensionality reduction, ultra-wideband, sensor radar. I. INTRODUCTION Counting people and things is important for crowd sensing and behavior analysis applications, including those related to the Internet-of-Things [1], smart environments [2], social networking [3], and surveillance [4]. For counting tasks, radiobased systems are preferred to image-based systems [5]-[7], especially when privacy, implementation costs, and obstructed line-of-sight represent important limitations. Among radiobased systems, device-free systems are often preferred to systems that rely on dedicated or personal devices [8]-[12]. Device-free systems are based on networks of sensor radars (SRs) that sense the wireless environment and detect targets from signal reflections (backscattering) [13]-[18]. The presence of obstacles and other scatterers (e.g., furniture, walls, and windows) leads to clutter and multipath propagation, which have detrimental effects on the detection performance. These phenomena are particularly severe in indoor environments, where the number of scatterers is large [19]-[21]. Conventional approaches for device-free counting via SRs rely on multi-target localization or tracking [22]-[24], where each SR estimates a set of metrics (e.g., ranges or angles) associated to a single detected target (namely, individualcentric approach). Typically, this approach has a complexity that grows exponentially with the number of targets due
