Research Progress of Diffusion Model in Crowd Activity Analysis

In crowded settings, crowd behavior analysis plays a crucial role in ensuring public safety. Although traditional generative models such as GANs and VAEs have been applied to tasks such as crowd anomaly detection and trajectory prediction, they still have many shortcomings in terms of multimodality, temporal consistency, and controllability. In recent years, diffusion models have emerged as a promising alternative due to their powerful generative capabilities. This paper provides a systematic review of the latest research progress in diffusion models for crowd behavior analysis, focusing on key tasks such as anomaly detection, crowd counting, trajectory prediction, and 3D human pose estimation. It first summarizes the basic principles of diffusion models, as well as representative advancements in noise scheduling, sampling strategies, and model architecture, while comparing their shortcomings with traditional generative models. Next, the paper proposes a new method classification system based on the degree of supervision: weakly supervised methods are categorized into three types based on the feature information utilized—global features, local features, and time series features; unsupervised methods are explored from the perspectives of reconstruction error, density modeling, and model optimization, with detailed analyses of the advantages and limitations of each method. Additionally, the paper summarizes the application potential of diffusion models in multiple crowd-related tasks, providing theoretical references and practical guidance for future research. This review aims to help researchers gain a deeper understanding of the role of diffusion models in crowd behavior analysis and to promote their further development and application in real-world scenarios.