A Statistical Method for People Counting in Crowded Environments Massimiliano Bozzoli 1. Hypercube S.r.L. v. di Strada Nuova s.n.c., 01030, Monterosi (VT), Italy m.bozzoli@hypercube.it Luigi Cinque , Enver Sangineto 2. Computer Science Department University of Rome “La Sapienza”, v. Salaria 113, 00198 Rome, Italy cinque@di.uniroma1.it, sangineto@di.uniroma1.it Abstract In this paper we present the results of a two-years research project on automatic people counting in public crowded environments. The aim of the proposed system is to estimate the number of people passing through a gate in a public area such as a metro or a railway station. The prob- lem is particularly challenging due to both the presence of crowd which makes it difficult the use of previous systems based on detection of isolated passengers and to the high level of statistic accuracy requested by traffic monitoring applications (error rate less then ). 1. Introduction Public transport companies (e.g., railway or bus compa- nies) need to estimate the number of passengers using their vehicles for marketing purposes. For instance, they need to know how many passengers transit in a given public station or get in and out of a bus or a train in different days of the year or in different hours of a day in order to optimize the route allocation and other services. Usually such statistics are obtained using ad hoc human personnel who manually counts the persons passing during a period of time (usu- ally, several hours long). The presence of crowd in these places makes it impossible to use other optic or mechanical means such as light beams or differential weight-based sys- tems which can work with only one people passing at once and usually cannot distinguish the passenger direction. As we will see later (Section 2), crowded situations are also the worst to deal with for most of the computer vision-based methods so far proposed for similar problems due to the difficulty in the segmentation process. Figure 2 shows two This work has been completely supported by the Hypercube company (www.hypercube.it). Info. to: Piergiorgio Scuriatti, p.scuriatti@hypercube.it., Tel.: +39-06-99806321. Corresponding author. Tel.: +39-06-49918358; fax: +39-06-8541842. typical (and realistically frequent) situations in which com- mon blob-based systems have difficulties: a crowd (Figure 2 (a)) and two touching people coming from opposite direc- tions (Figure 2 (b)). Another major difficulty of the problem is the high accu- racy ( ) requested to the statistic data to be valid. The first prototype version of the system we propose was able to reach this target value. In this paper we present the results of a two years project sponsored by a private consortium interested in the automa- tion of the people counting process. The research project has been supported and developed by the Hypercube com- pany and tested on about three hours of traffic scenes ac- quired in a Roman railway station in different lighting con- ditions. The image examples shown in this article have been extracted from these test video sequences. 2. Related Work Previous works on this subject usually deal with a sparse passenger frequency or are based on the explicit or implicit assumption of a one-to-one matching between humans in the observed scene and foreground blobs (i.e., connected sets of pixels detected by means of various segmentation techniques). For instance, in the system proposed by Albiol and colleagues [1], built for the Spanish Railway Company with the same objective of the present paper, the camera is mounted on the train’s door ceiling and people getting in and out of the train are assumed to be separated the one from the others by a fraction of time necessary to the algo- rithm to segment each person in a different blob. A blob is segmented looking for a strong vertical gradient value which indicates the train’s step boarder visibility and hence the absence of passing passengers. With this segmentation assumption the system achieves very good experimental re- sults (error rate less than ) but in an unconstrained en- vironment or with larger doors in which people can walk together it is unlikely to work correctly. A skin-colored ellipse representing the passenger’s head