Analyzing human crowds is becoming an important issue in video surveillance and one challenging task is to detect group-level crowd due to their non-rigid shapes nature. This study presents a novel method which synergistically combining two state-of-the-art methodologies to identify groups in crowds. The first is the ability to track crowd trajectories using particle video technology and the second is a new class of novelty clustering algorithms based on spectral analysis of graph. Simultaneity, the social science principle of human collective behavior, such as the similarity of location, velocity, appearance, is also inspired to cluster crowd trajectories. Experimental results demonstrate that our method is effective in tracking and identifying group-level crowds for public surveillance videos.
展开▼
