Content-Based Video Copy Detection Scheme Using Motion Activity and Acoustic Features

Abstract

This paper proposes a new Content-Based video Copy Detection (CBCD) framework, which employs two distinct features namely, motion activity and audio spectral descriptors for detecting video copies, when compared to the conventional uni-feature oriented methods. This article focuses mainly on the extraction and integration of robust fingerprints due to their critical role in detection performance. To achieve robust detection, the proposed framework integrates four stages: 1) Computing motion activity and spectral descriptive words; 2) Generating compact video fingerprints using clustering technique; 3) Performing pruned similarity search to speed up the matching task; 4) Fusing the resultant similarity scores to obtain the final detection results. Experiments on TRECVID-2009 dataset demonstrate that, the proposed method improves the detection accuracy by 33.79% compared to the referencemethods. The results also prove the robustness of the proposed framework against different transformations such as fast forward, noise, cropping, picture-inpicture and mp3 compression.