Video analysis and abstraction in the compressed domain.

摘要

The objective of this work is to develop a fast system for and can create images and short video clips according to a multi-resolution scheme for interactive video browsing. The proposed approach is camera event-based and is designed around a simple, fast, and reliable algorithm. Specifically, to reduce the high computational load, the measurements used to make decisions for key-frame extraction come directly from the compressed domain. First, shot segmentation and camera movement classification are performed. Next, a simple key-frame extraction algorithm selects the representative frames that are clustered according to their visual similarities using a new and fast clustering algorithm that combines the SVD and k-means algorithms. Finally, the clustered frames are assembled into a video summary using a hierarchical scheme. In addition, a new method for detecting human another algorithm, it is demonstrated that the proposed approach is fast and produces an effective video summary. An average overall processing time that is 9 times faster than real-time on a Pentium II, 400 MHz PC has been achieved for 352 × 240 video sequences.