Matching Distorted Video Clips to a Reference Video Database

The objective of this study is to develop and validate video matching algorithms that satisfy the quality, performance, scalability, and stability requirements necessary for a commercial video matching service. The study assumes that short, distorted video clips are matched against a large reference video database containing at least 10,000 hours of video. The literature review revealed few studies that came close to the reference video database size, match accuracy and transaction rate requirements used in this study. Very few large publicly available datasets for partial video copy detection are available making comparisons with other studies difficult. Synthetic and real-world video datasets were constructed to assess the performance of the various video matching algorithms used. Both machine learning and non-machine learning algorithms are investigated in this study. The matching algorithms extract and compare frame-level feature vectors from the videos to determine a match score. A variety of techniques are then used to determine whether a particular video clip matches a sequence in the reference video database. Surprisingly, the non-machine learning algorithms performed better than the machine learning approaches with the overall study objectives achieved by a hybrid algorithm. This hybrid algorithm was found to out-perform the throughput and match accuracy of other partial video copy detection algorithms found in the literature.