Skip to Main content Skip to Navigation

Towards camcorder recording robust video fingerprinting

Abstract : Sine qua non component of multimedia content distribution on the Internet, video fingerprinting techniques allow the identification of content based on digital signatures(fingerprints) computed from the content itself. The signatures have to be invariant to content transformations like filtering, compression, geometric modifications, and spatial-temporal sub-sampling/cropping. In practice, all these transformations are non-linearly combined by the live camcorder recording use case.The state-of-the-art limitations for video fingerprinting can be identified at three levels: (1) the uniqueness of the fingerprint is solely dealt with by heuristic procedures; (2) the fingerprinting matching is not constructed on a mathematical ground, thus resulting in lack of robustness to live camcorder recording distortions; (3) very few, if any, full scalable mono-modal methods exist.The main contribution of the present thesis is to specify, design, implement and validate a new video fingerprinting method, TrackART, able to overcome these limitations. In order to ensure a unique and mathematical representation of the video content, the fingerprint is represented by a set of wavelet coefficients. In order to grant the fingerprints robustness to the mundane or malicious distortions which appear practical use-cases, the fingerprint matching is based on a repeated Rho test on correlation. In order to make the method efficient in the case of large scale databases, a localization algorithm based on a bag of visual words representation (Sivic and Zisserman, 2003) is employed. An additional synchronization mechanism able to address the time-variants distortions induced by live camcorder recording was also designed.The TrackART method was validated in industrial partnership with professional players in cinematography special effects (Mikros Image) and with the French Cinematography Authority (CST - Commision Supérieure Technique de l'Image et du Son). The reference video database consists of 14 hours of video content. The query dataset consists in 25 hours of replica content obtained by applying nine types of distortions on a third of the reference video content. The performances of the TrackART method have been objectively assessed in the context of live camcorder recording: the probability of false alarm lower than 16 10-6, the probability of missed detection lower than 0.041, precision and recall equal to 0.93. These results represent an advancement compared to the state of the art which does not exhibit any video fingerprinting method robust to live camcorder recording and validate a first proof of concept for the developed statistical methodology.
Document type :
Complete list of metadata

Cited literature [158 references]  Display  Hide  Download
Contributor : ABES STAR :  Contact
Submitted on : Thursday, October 10, 2013 - 1:42:09 PM
Last modification on : Wednesday, November 17, 2021 - 12:30:58 PM
Long-term archiving on: : Saturday, January 11, 2014 - 4:20:33 AM


Version validated by the jury (STAR)


  • HAL Id : pastel-00871762, version 1


Adriana Garboan. Towards camcorder recording robust video fingerprinting. Other [cs.OH]. Ecole Nationale Supérieure des Mines de Paris, 2012. English. ⟨NNT : 2012ENMP0097⟩. ⟨pastel-00871762⟩



Record views


Files downloads