Graph-based correlated topic model for trajectory clustering in crowded videos

This paper presents a graph-based correlated topic model (GCTM) to analyse various motion patterns by trajectory clustering in a highly cluttered and crowded environment. Unlike the existing methods that address trajectory clustering and crowd motion modelling using local motion features such as optical flow, it builds on trajectory segments extracted from crowded scenes. Correlated topic models have been previously applied to handle mid-level features learning in crowded scenes. However it depends on scene priors in the learning process. GCTM addresses this issue by using a spatio-temporal graph and manifold-based clustering as initialization and iterative statistical inference as optimization. The output of GCTM is mid-level features used later as an input to the final step that generates trajectory clusters. Experiments on two different datasets show the effectiveness of the approach in trajectory clustering and crowd motion modelling.