A Review of Deep Learning Techniques for Crowd Behavior Analysis

In today's scenario, there are frequent events (viz. political rallies, live concerts, strikes, sports meet) occur in which many people gather to participate in the event. In crowded areas possibility of occurrence of suspicious activities and violence automatically increases. To attenuate these issues, it is very important to develop automated systems that can detect anomalies in the diverse and complex outdoor environments, especially at public places to ensure safety and to avoid crowd disasters such as human stampede, mob lynching, and riots. Crowd management may help to circumvent the crowd disaster and ensure public safety at places like temples, railway stations, airports, bus terminals, religious functions, political rallies etc. We need a visual surveillance that can automatically detect abnormality in crowd behavior so that the relevant action can be taken to prevent any public casualty. Basic steps required for crowd analysis is density estimation and crowd counting, object recognition, tracking, and anomaly detection in crowded scene. Here, we have systematically reviewed and compared different methods that are being used for crowd analysis. The comparative analysis of existing methods has been presented on different available datasets and taxonomies are also compared. We also propose our own taxonomy for crowd analysis and datasets.

Automated summary

In today's crowded events, the possibility of suspicious activities and violence increases. Developing automated systems for crowd analysis is important to ensure public safety and prevent disasters. This analysis involves density estimation, crowd counting, object recognition, tracking, and anomaly detection. Comparing different methods and creating taxonomies for crowd analysis are also important for effective implementation.