Summary: | 碩士 === 國立臺灣科技大學 === 電子工程系 === 103 === Nowadays, the video surveillance systems have become popular, and are deployed in every corner of our environment, such as in the school, airport, street, etc. It generates huge data consistantly. Thus, browsing and searching for a specific event from this database becomes challenging. To reduce the efforts by manually searching targets, the video synopsis technique was proposed to efficiently capture the dynamic behavior of a specific object with limited time period. The video synopsis method provides a condensed video by removing the spatial or temporal redundancies. This technique maintains and records all the activities completely in the original video. Some former methods have been proposed, yet they require long computational time. In addition, the blanking effect is accompanied. To overcome the above issues, a new efficient trajectory-based video synopsis is proposed in this thesis. It consists of four parts, including 1) trajectory-based object classification to keep the tubes continually and avoid the blanking effect on the synopsis video; 2) Minimum Overlap (MO) algorithm for the objects tubes temporal position in synopsis video, 3) Global Temporal Shifting (GTS) processes to make the tubes with the temporal-domain flexibility, and 4) Attributes search for a efficient way to enable a user locating specific events precisely and efficiently. As demonstrated in the experimental result, the proposed video synopsis effectively produces smooth synopsis videos without blanking, jumping, and ghosting issues.
|