Periodic motion detection and segmentation via approximate sequence alignment

摘要: A method for detecting and segmenting periodic motion is presented. We exploit periodicity as a cue detect in complex scenes where common methods segmentation are likely to fail. note that detection can be seen an approximate case of sequence alignment image matched itself over one or more periods time. To use this observation, we first consider two video sequences obtained by independently moving cameras. Under assumption constant translation, the fundamental matrices homographies shown time-linear matrix functions. These dynamic quantities estimated matching corresponding space-time points with similar local shape. For motion, match across develop RANSAC procedure simultaneously estimate period geometric transformations between views. Using method, demonstrate human nonrigid backgrounds, camera parallax.