| With the explosive growth of video content on the Internet and for personal storage, an efficient and robust example clip-based copy detection solution is needed to support applications like query by clip, repeat clip detection, and copyright violation detection. In content-based copy detection applications, video content representations are essential as they determine how well the desired video separate from unrelated videos in the database. There are different features that have been utilized for this purpose, such as color histogram, shape, texture, motion, etc. In this paper, multiple algorithms are proposed for content-based video retrieval. When using global features, video sequences are represented as temporal trajectories via scaling and linear transformation of the frame luminance field. An appropriate lower dimensional subspace is identified for video trajectory indexing via Principal Component Analysis (PCA) or Random Projections. Indexing is very important to the retrieval as it allows quick searching through the database and eliminating unrelated video clips. We propose two different indexing structures for this purpose. One is the well-known kd-tree indexing scheme. The other one is a novel idea called LUFT-tree. We also develop a trace geometry-matching algorithm for retrieval based on average projection distance. Simulation results demonstrate the high accuracy and retrieval speed achieved by the proposed solution. Sparsity is a recent popular concept for data acquisition and compression. Due to its interesting properties related to the video retrieval application, we propose employing its framework for our purpose. Simulation produced very encouraging results. We also propose two algorithms that use local features. One uses Fisher Vector to aggregate local descriptors (SIFT or BRISK) to an image-level signature. The other uses multi-basis LPP to compress SIFT descriptors. It showed that the proposed algorithms using global features can achieve retrieval accuracy with some robustness to occlusion and corruptions, while the proposed algorithms using local features perform well even when the videos are under different types of transformation attacks. |
