Dynamic modeling of streaklines for motion pattern analysis in video

Natural videos consist of multiple objects interacting with each other in complex ways. Analysis of such videos is a very challenging problem. The underlying patterns of motion contain valuable information about the activities and hence, it is necessary to devise efficient mechanisms for extraction and analysis of these patterns. Tracking is often infeasible in these scenarios. Recently, flow-based modeling methods based on the concept of streaklines in fluid mechanics have been proposed to deal with videos of crowds. In this paper, we show how the streaklines representation can be combined with dynamical systems modeling approaches in order to analyze the motion patterns in a wide range of natural videos. This combination provides a powerful tool to identify similar segments in a video that exhibit similar motion patterns. We show results on challenging datasets by demonstrating unsupervised clustering of similar motion patterns (which often represent similar activity patterns) and detection of changes in the video.