Binkey: a system for video content analysis 'on the fly'
Résumé
A system for automatic and fast structuring of video sequences allowing detection of shot changes, dominant motion estimation, and extraction of areas of interest is proposed. It is based on the analysis of a 1-D representation of video frames by a Mojette transform, a discrete version of the Radon transform. Motion is estimated by analysis of the projected motion in the transform domain. The proposed method uses statistical techniques based on cross-correlation measurement and outliers rejection which ensure its general robustness, specifically in case of mobile object. Results of this computation are then used for the detection of shot change and for the localization of areas of interest in the image plane. The analysis of a 1-D representation of video frames rather than the usual 2-D image reduces the complexity by itself and gives the possibility to video partitioning and object extraction 'on the fly'.