Scene duplicate detection from videos based on trajectories of feature points

• Authors: Shin'ichi Satoh;Masao Takimoto;Jun Adachi
• Venue: Proceedings of the international workshop on Workshop on multimedia information retrieval
• Year: 2007

Recently near duplicate shot detection attracts researchers attention. There are several promising applications of near duplicate detection, especially when applied to broadcast video streams. However, currently studied near duplicates are limited to: a) The same video material (footage) used several times in different programs (strict near duplicate), and b) Less strict near duplicates such as footages of the same objects or the same background (object duplicate). In this paper, we propose a method to detect scene duplicates, another type of near duplicate. This type of near duplicate is composed of different footages taking the same scene, the same event, at the same time, but from the different viewpoints, e.g., by the different cameras, and possibly with temporal offsets. This type of near duplicate is particularly useful to identify the same event reported in the different programs and especially by the different broadcast stations. To handle this, we employ matching of temporal pattern of discontinuities obtained from trajectories of feature points. To detect discontinuities and to match trajectories, we used inconsistency. The method is accelerated by two-stage approach: the first stage is filtering by using temporal discontinuity patterns, and the second is precise matching by normalized cross correlation between inconsistency sequences of trajectories. The first stage is further accelerated by using interval histograms. Its performance is demonstrated with actual broadcasted videos from five channels and three hours from each channel, in total 15 hours.