Activity analysis in complicated scenes using DFT coefficients of particle trajectories

摘要

Modelling activities in crowded scenes is very challengingudas object tracking is not robust in complicated scenesudand optical flow does not capture long range motion. Weudpropose a novel approach to analyse activities in crowdedudscenes using a “bag of particle trajectories”. Particle trajectoriesudare extracted from foreground regions within shortudvideo clips using particle video, which estimates long rangeudmotion in contrast to optical flow which is only concernedudwith inter-frame motion. Our applications include temporaludvideo segmentation and anomaly detection, and we performudour evaluation on several real-world datasets containingudcomplicated scenes. We show that our approaches achieveudstate-of-the-art performance for both tasks.