Sensing, Analysing and Simulating Heterogeneous Unstructured Crowds in Mass Religious Gatherings

Abstract

Mass gatherings have witnessed several incidents such as crowd crushes, stampedes, panic rush, etc. In the last 10 years, nearly 49 major incidents have led to more than 4000 casualties and similar number of injuries. In the current scenario in India, crowd managers mostly make decisions based primarily on manual observation and prior experience, which are mostly subjective. Existing pedestrian sensing and simulation models are majorly developed outside India. They do not capture the varied pedestrian characteristics, behaviour, and constraints observed in gatherings in Indian scenarios. One of the key reasons is the need for datasets for such pedestrian sensing models that capture the heterogeneous crowds observed in mass gatherings in India. This thesis focuses on developing pedestrian sensing and simulation models to aid decision support for crowd management in mass religious gatherings. The data was collected during the Kumbh Mela 2016 in Ujjain, using various surveillance devices for a month. In Kumbh Mela, pilgrims arrive in millions comprising people from all sects, rural, urban, foreign, and saints. This makes Kumbh Mela an appropriate choice for the study, as most of the scenarios observed in Kumbh Mela can also be found in other religious gatherings in India. To develop pedestrian sensing models for situations observed in Kumbh Mela, datasets are first developed to capture complexity. Second, pedestrian sensing models are developed at three levels: individuals, groups and crowds. At the Individual level, (i) pedestrian counting under poor lighting conditions and (ii) pedestrian category-wise counting for the heterogeneous crowds was done. In the first part, a detection-based count model and a denoising filter were used to estimate pedestrian density. In the second part, data from a heterogeneous dense crowd was used to count males, females, and pedestrians carrying head luggage. A novel unsupervised group detection and tracking model is developed at the group level for medium pedestrian density. The study also includes the method for counting group splits and group merges and new techniques for assessing and evaluating group sensing models. At the crowd level, a (i) semi-supervised crowd flow estimation model and an (ii) unsupervised crowd entropy estimation model are developed. In the first part, the model combines deep learning and image processing techniques for estimating the crowd flow. In the second part, image processing techniques estimate the crowd entropy. Other macroscopic parameters and entropy are then modelled to assess the potential crowd risk. Third, a pedestrian sensing tool, comprising a pedestrian detection and tracking model, is developed to estimate crowd parameters. Finally, certain pedestrian behaviours observed during Kumbh Mela are modelled and simulated. Two pedestrian behaviours observed in Kumbh Mela, (i) Serpentine group behaviour and (ii) Lane changing behaviour in single file channel, are simulated. In the case of Serpentine group behaviour, groups form a serpent-like structure and attempt to sneak through the crowd to reach their destination early. This behaviour was modelled by modifying the social force model and was simulated for different geometry conditions. In the case of a single file channel, pedestrians form two virtual lanes and make lane changes to reach their destinations early. The behaviour is simulated under different scenarios and crowd control strategies. The developed simulation models assist in better understanding pedestrian behaviour under other circumstances. The study provides valuable insights for pre-event planning and live crowd monitoring that could help authorities manage crowds better.