基于时空轨迹行特征的视频拷贝检测方法.docVIP

基于时空轨迹行为特征的视频拷贝检测方法 吴 潇1,2 李锦涛1 唐胜1 郭俊波1 1 (中国科学院计算技术研究所前瞻研究实验室 北京 100190) 2 (中国科学院研究生院 北京 100049) (wuxiao@) Video Copy Detection Based on Spatio-Temporal Trajectory Behavior Feature Wu Xiao 1, 2, Li Jintao 1, Tang Sheng1, and Guo Junbo 1 1(Center for Advanced Computing Research, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190) 2(Graduate University of the Chinese Academy of Sciences, Beijing 100049) Abstract Large scale video copy detection is to detect copied segments of provided video content from large video databases. This application requires compact feature which is insensitive to various visual copy changes. However, traditional image features are prone to spatial changes such as color and texture transformations. The reason is the copied versions have large image transformations including global quality decrease and local visual distortions, which dramatically change the distribution of visual features. Consequently, previous methods based on histogram features and ordinal measures fail in copy detection. To solve this problem, this paper proposes the use of invariant visual features based on keypoint trajectory behavior. Instead of using the spatial cues, the proposed approach models the temporal information as robust features. Temporal cues are quantified based on keypoint trajectories which are insensitive to strong changes. Then videos are represented by spatio-temporal features which are more robust to copy changes. Bag of Trajectory (BoT) technical is adopted for fast pattern matching in large database. The experiments show that spatio-temporal trajectory features are robust to various visual changes, including image blur, scale and ratio changes, and minor frame rate change. Compared with the state-of-art scheme using ordinal measure, the proposed algorithm with lower cost presents better accuracy. Key words video copy detection; spatio-temporal trajectory; local feature point detection; matching based on tempora