Video analysis for a detailed reconstruction of individual and crowd dynamics

Crowd density estimation and crowd dynamics analysis are critical tasks in modern urban management, event planning, and public safety. This thesis explores a range of computer vision methodologies to estimate crowd density in the unique context of Venice's crowded public spaces. The study implements and evaluates several techniques, including the state-of-the-art YOLOv8 object detection model, a custom U-Net for crowd segmentation, a traditional background subtraction algorithm, and a novel approach based on trajectory tracking. Each method is evaluated in terms of accuracy and computational efficiency on both GPU-powered systems and resource-constrained devices like the Raspberry Pi 5. While YOLOv8 delivers high accuracy in low-density crowds, it struggles in dense settings and demands substantial computational resources. In contrast, the U-Net-based model shows reliable performance across both low and high-density scenarios, though its counting accuracy can be imprecise. The traditional background subtraction approach excels in real-time processing on low-power devices but its output tends to be less precise and harder to interpret. The novel tracking-based approach offers a promising alternative for estimating crowd density by analyzing trajectory patterns, though quantitative evaluation wasn't possible and more research need to be conducted. The results provide a comparative framework to guide the selection of crowd analysis techniques based on the density of the crowd and the available computational resources. This work not only contributes to the research in crowd analysis and computer vision, but also lays the foundation for future research in real-time crowd management and analysis.