Poster Abstract: Crowd Crush Detection in Large Mass Gatherings via Federated Learning Across Multicamera Environment

Abstract

Crowd crush in mass gatherings during large events are more frequent nowadays. It may get threatening quickly and even cause the death of many. Single camera surveillance is not efficient in large events spreading over wide areas. In multicamera surveillance environment, the centralized decision may not be favourable due to solitude and safety concerns. Moreover, existing neural networks need a large amount of training data for accurate detection. This paper proposes a low-cost system that uses a federated learning setup with light-weight model to reduce overheads during training and protect crowd privacy. We introduce a new loss function, denoted by contrastive focal loss, to reduce false positives and organize overcrowded regions in minimal time. We collect crowd crush and stampede videos to create a new annotated dataset, named as CrowdStampede. We achieve good results under different data distribution settings. © 2022 ACM.