Skip to Main content Skip to Navigation
Theses

Indexation vidéo par l'analyse de codage

Abstract : This thesis work concerns indexation, normalized by MPEG7, of video sequences. From a MPEG1-2 stream, or from any other codec based on movement prediction and DCT, without decompressing it completely, we exploit the analysis carried out during the encoding process. This way, unsupervised and in quasi real-time, we provide a method to estimate the camera movement as well as moving objects extraction. As far as camera movement estimation is concerned, we use motion vectors included in the stream. Studying the error images allows us to assess its relevance. In order to detect moving objects, we first segment the sequence into uniform color zones directly on the DCT coefficients. We establish a colorimetric distance, not only between two neighbouring pixels in the same image, but also in two successive images, which allows to define a three-dimensional zone. In order to provide a more accurate segmentation, and to regularize the contours on every image, we use B-splines. Every candidate object is distorted by the presence of all its neighbors, based on a color potential. This allows iteratively to eliminate zones which were excessively reduced. By combining camera movements, prediction vectors and 2D+t color zones, we create an adaptative fusion in order to obtain a good representation of objects, and thus of their monitoring.
Document type :
Theses
Complete list of metadata

Cited literature [117 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00214113
Contributor : Estelle Nivault <>
Submitted on : Wednesday, January 23, 2008 - 1:49:09 PM
Last modification on : Wednesday, October 14, 2020 - 4:23:55 AM
Long-term archiving on: : Thursday, September 27, 2012 - 5:39:22 PM

Identifiers

  • HAL Id : tel-00214113, version 1

Collections

Citation

Lionel Brunel. Indexation vidéo par l'analyse de codage. Automatique / Robotique. Université Nice Sophia Antipolis, 2004. Français. ⟨tel-00214113⟩

Share

Metrics

Record views

324

Files downloads

1144