Abstract
In this thesis, some efficient methods that can explore motion activity features from motion vectors for inquiring video clip on MPEG compressed domain are proposed. The reasons that we are interested in utilizing the compressed-domain motion activity features are as follows. First, much of the multimedia content available today is in compressed form already. Using compressed-domain features directly makes it possible to construct effective video indexing and retrieval systems for real-time applications. Next, motion activity features are easily to be extracted from the existing motion vectors without the need of computation of decompression as well as the delay latency induced. We analyze and utilize the motion activity features we extracted to build up a system for inquiring video clip by example. With the continuing explosion of multimedia information in today's society, searching for information of interest is becoming increasingly hard. As it is not straightforward to describe video content in keywords, searching video solely on text information has its limitations. We deal with this problem through the technique of query by example. First a matching algorithm is designated to evaluate the similarity between the existing video segment and the query one derived from the information of their corresponding motion activity features. Then, dynamic programming technique is employed to search for the most proximity video segment in the lengthy question video with a reasonable computational cost. Finally, some applications are given and discussed to demonstrate the potential usage of the proposed techniques.