Using Entropy of Social Media Location Data for the Detection of Crowd Dynamics Anomalies

Abstract

Evidence of something unusual happening in urban areas can be collected from different data sources, such as police officers, cameras, or specialized physical infrastructures. In this paper, we propose using geotagged posts on location-based social networks (LBSNs) to detect crowd dynamics anomalies automatically as evidence of a potential unusual event. To this end, we use the Instagram API media/search endpoint to collect the location of the pictures posted by Instagram users in a given area periodically. The collected locations are summarized by their centroid. The novelty of our work relies on using the entropy of the sequence of centroid locations in order to detect abnormal patterns in the city. The proposal is tested on a data set collected from Instagram during seven months in New York City and validated with another data set from Manchester. The results have also been compared with an alternative approach, a training phase plus a ranking of outliers. The main conclusion is that the entropy algorithm succeeds inn finding abnormal events without the need for a training phase, being able to dynamically adapt to changes in crowd behavior.