Deep Fundamental Diagram Network for Fast Pedestrian Dynamics Estimation

How to effectively guide occupants to use different evacuation routes under fire situations is the key to improving fire safety and ensuring successful evacuation. Evacuation analysis for fire safety in surveillance videos plays a crucial role in understanding and mitigating risks. The fundamental diagram of pedestrian flow, which illustrates the relationship between pedestrian velocity and crowd density, is a valuable tool for analyzing evacuation dynamics and enhancing fire safety measures. Traditional methods rely on trajectory files obtained from manually tracking each pedestrian in video recordings to construct fundamental diagrams. However, these methods have limitations in accurately representing crowd density and cannot provide real-time analysis, making them unsuitable for surveillance camera analysis in fire safety scenarios. To address this challenge, we propose a novel convolutional neural network-based framework called the deep fundamental diagram network, which is specifically designed for fire safety applications. This framework consists of two modules: the multi-level dilated convolutional neural network (MLD-Net) and the optical flow module. The MLD-Net learns the mapping relationship between input images and density maps, enabling accurate estimation of pedestrian density. Simultaneously, the optical flow module calculates pedestrian movement speed, providing crucial information for evacuation planning. By aligning the density map with the speed map, the fundamental diagram is derived, which aids in understanding evacuation dynamics. The experimental results demonstrate that our method achieves good consistency with traditional approaches while significantly reducing the computational time. Additionally, our framework enables anomaly detection and pedestrian line counting, further enhancing fire safety measures. This work is expected to have good prospects in the fields of fire safety, evacuation dynamics analysis, and real-time crowd analysis systems for fire situations.