Human crowd detection for physical sensing assisted geo-social multimedia mining

Real-time context monitoring in public indoor space is a key technology for cyber-physical systems. Recent research studies have shown that analysis of geo-tagged multimedia data on social networking websites enables accurate detection and localization of “events” in real world, which attract attention of a number of people (e.g., earthquakes). Although the similar idea could also be applied to local event detection in indoor space, the spatial locality of the events makes it a much more challenging task. Since the number of people who observe the same event can be severely limited compared to global events like earthquakes, it is hard to distinguish the meaningful posts that help event detection from a pile of irrelevant data. To address the issue, we propose a hybrid event detection mechanism, which effectively combines physical crowd behavior sensing using laser range scanners (LRSs) and social multimedia mining. The basic idea behind our approach is that meaningful events (e.g., an interesting demo at a specific booth in an exhibition hall) usually attract attention of the surrounding pedestrians, forming a human crowd. We detect such a sign of events (i.e., human crowds) in each area by pedestrian tracking using a small number of LRS sensors, and filter the posts based on the associated geo-tags. A major challenge of this approach is that the tracking system often underestimates the number of pedestrians in such crowded regions due to occlusion of measurement signals, causing frequent misdetection of human crowds. To cope with the problem, we build an empirical model that identifies relationship between actual crowd density and the number of pedestrians who are captured by the LRS sensors, with which we correct errors in crowd density estimation. Through simulations, we show that our system can accurately detect locations of human crowds even under severe occlusion.