Video Copy Detection Based on Bag of Trajectory and Two-Level Approximate Sequence

Abstract—We present a video copy detection system that detects video copy segments based on the task settings and dataset in TRECVID 2010. Contributions of this work are two-fold. First, we extract feature-based trajectories from videos, and then model trajectories by a bag of word model. This representation effectively describes information of object movement, and is robust to various visual transformations. Second, to find locally optimal matching between the query video and videos in database, we conduct approximate sequence matching by first finding the longest common subsequence, followed by localizing the max-sum segment. We compare our method with a watershed-based approach and demonstrate the effectiveness and robustness of the proposed method. Keywords-video copy detection; bag of trajectory; the maximum-sum segment algorithm; approximate sequence matching I