| Summary: | 碩士 === 國立臺灣科技大學 === 電機工程系 === 102 === Video synopsis is utilized to provide a condensed video without spatial and temporal redundancies, while maintaining all the activities in the original video. Existing video synopsis methods presume that the object tubes (three dimension space time volume) are perfectly retrieved from the original video, yet it is not the case in practical applications due to the object detection/tracking failure or object collision. Another problem is the energy minimization based method often suffers from heavily computational cost, which thus prolongs the processing time. In this thesis, a new video synopsis scheme is proposed to solve the above issues.
The main contributions of this study are as follows: 1) A highly efficient background subtraction scheme with dynamic edge thresholding is proposed to provide an artifact-free video synopsis result, and preserve most of the non-specific objects; (2) the processing speed of the overall video synopsis is even faster than real-time requirement to enable the multi-channel scenario; (3) overcome the problems induced by the preprocessing such as background subtraction and object tracking, and create fluent synopsis video without decreasing processing speed, and (4) overcome the energy minimization problem in former works with the object correlation clustering and properity weighting function. As documented in the experimental results, the proposed scheme provides promising performance, and can be a very competitive candidate in the practical video synopsis application.
|
|---|
