Ensemble learning for large-scale crowd flow prediction

Abstract

The Kumbh Mela festival is the largest mass gathering in the world that is celebrated every three years. In 2016, it attracted over 70 million people to Ujjain, India. The Mahakal temple is the “heart” of the festival that attracts a huge number of pilgrims and needs to accommodate with massive crowds. These types of events pose significant safety challenges as large-scale mass gatherings are often associated with risks such as crowd crushes. There have been a number of serious incidents documented in recent history such as the Hajj crush at Mina, Mecca, Saudi Arabia (2006 and 2015), the Lame Horse crush during a fire at Perm, Russia (2009), the Love Parade disaster at Duisburg, Germany (2010), and the Kumbh Mela stampede at Allahabad, Uttar Pradesh, India (2013) to name a few. Safety assurance at events of such tremendous size is closely connected with crowd control and understanding the general behaviour of the crowd. One of the basic challenges in understanding crowd dynamics is being able to predict crowd flows at a particular location based on past/present flows from another location. There are several existing methods and models used to predict and manage crowd flow. In this paper, we introduce a novel method for short-term crowd flow prediction and show that it decreases the prediction error by 13% as compared to existing methods. The model is based on ensemble learning where we demonstrate that a combination of complementary methods with different a-priory assumptions can create better estimations. Utilizing a unique data set derived from CCTV camera recordings of pilgrims that we collected during the Kumbh Mela 2016 festival, we tested different methods from artificial intelligence and computational modelling, such as simple shift of time-series (time-shift), agent-based modelling, machine learning methods, and show how combinations of these different methods as an ensemble provide synergy to obtain better predictions. Our results demonstrate that agent-based modelling, when combined with other models, provides better predictive power especially in complex scenarios. These results point to something fundamental about the information contained within and generated by these methods. We anticipate that our research could be a starting point for further research of informational synergetic aspects of models and predictors.