Background: Very dense crowds that exceed three people per square metre present many challenges in computer vision for efficiently measuring quantities such as density and pedestrian trajectories. An accurate characterisation of such dense crowds can improve existing models and help to develop better strategies for mitigating crowd disasters. Pedestrian models used for tracking are often based on assumptions which are no longer valid in highly dense crowds, e.g. absence of occlusions. Recently, multiple camera systems with partially overlapping fields of view have been shown to offer critical advantages over other non-overlapping schemes. Objectives: We focus on overcoming the non-invasive aspect of the camera calibration, imposed by real crowded environments needed in order to accurately segment individuals. We will also underline the interdisciplinary challenges related to computer vision and real-time processing. Methods: Although there is an important amount of work devoted to multiple camera calibration, the automation of the process in specific environments remains challenging. In dense crowds, such as in Mecca, access to the site for calibration purposes or for adding support visual features is impossible. Moreover, the supervising role of the user for calibration scenarios is very important, and cumbersome for large camera networks. We investigate a light solution based on a coarse-to-fine estimation of the camera positions using both static and dynamic features. This highlights the necessary tradeoff between the crowd coverage, the purpose of the experiment, and the static feature distribution which is required to register the camera system properly. A more practical aspect that we underline is related to the importance of accurate time synchronization within the system in the presence of a dynamic scene. Results: We present a pilot study of the above scheme conducted at Regents Park Mosque in London on Friday when the mosque is particularly crowded. We have set up a distributed system of accurately synchronized Firewire cameras, acquiring high-resolution data at 8Hz. We will also aim to present some preliminary single camera studies of crowd flow using real-world data from the Muslim Hajj pilgrimage from 2011.


Article metrics loading...

Loading full text...

Full text loading...

This is a required field
Please enter a valid email address
Approval was a Success
Invalid data
An Error Occurred
Approval was partially successful, following selected items could not be processed due to error