Cluster behavior recognition technology

The scene complexity and derived social risks of cluster behavior are far greater than individual behavior. Cluster behavior recognition has emerged as a key research area in both computer vision and public safety. Methods in this domain have seen continuous advancements, with steady improvements in recognition speed and accuracy. This paper provides a comprehensive review of cluster behavior recognition technology as documented in recent literature. Based on feature sources, recognition methods are categorized into manually designed approaches and machine learning-based models. Manually designed approaches include methods focused on overall cluster characteristics and those analyzing individual behaviors within groups. Machine learning models are further subdivided into non-interaction modeling, global member interaction modeling, and key individual interaction modeling. Key technologies in machine learning for cluster behavior recognition are also examined, with emphasis on feature extraction techniques like Long Short-Term Memory Networks, 3D Convolutional Neural Networks, and Inflated 3D Convolutional Neural Networks, as well as interaction modeling approaches using Graph Convolutional Networks and Graph Attention Networks. Finally, this paper discusses the problems that need to be solved urgently and prospects the future development trend of cluster behavior recognition.