Robust video copy detection in large-scale TV streams using local features and CFAR based threshold

In this paper, a robust video copy detection system is proposed for the broadcast TV stream that is a challenging task in the wild. The proposed system extracts the signatures of video frames with Speeded-Up Robust Features (SURF) keypoints, which are described by Oriented Fast and Rotated Brief (ORB) descriptors for compact and efficient representation, in order to match the parts of the given reference video with the query video in interest. The resulting signal representing the similarity scores of the reference video parts with the query is then adaptively thresholded by Constant False Alarm Rate (CFAR) approach, which improves the video copy detection accuracy by exposing the strong peaks (i.e., copy locations) while discarding weak peaks/no peaks (i.e., non-copy locations), with dynamic setting of the false alarm probability. Extensive experiments on the publicly available MUSCLE-VCD-2007 Dataset and the BILGEM Video Broadcast Dataset 2015 (BVBD 2015) show that the proposed system achieves precision values of 100.0% and 91.8%, recall values of 90.0% and 99.3% respectively by proving the robustness of the system for the copy detection task in both broadcast and web video streams.