Summarization of Multi-view Surveillance Videos by an Object-Based Key Frame Extraction Method
 
 
 
Yao-Ching Huang (黃耀慶)
中華民國 100 年 6 月
輔仁大學---智慧型系統實驗室---雲端視覺組
 
 
 
 
Abstract
    Video summarization is an important technique which has been an interested subject in many research fields which generates a short summary of a video for the presentation to users with browsing and navigation. Multi-view development is also beneficial to video surveillance, since the vast public security area installed a lot of cameras need to filter of huge non-important information. In this paper, we propose a multi-view video summarization approach that extracts semantic-level key frames by object information from multiple cameras. Our main goal is to avoid the redundant key frames with multi-view videos that the dominant camera selection presented to decentralize key frame extraction approach. The proposed approach is a new formulation which integrates camera selection algorithm into key frame extraction for optimization.
    This proposed approach has been verified by large amounts video dataset that include different surveillance scenes, and comparing with other camera selection method. This method proved by experiments not only can extract representative key frames but also reduce redundant key frames in multi-view videos.
 

 
 
Method
    In this paper, we propose a multi-view video summarization approach that extracts semantic-level key frames by object information from multiple cameras with dense or sparse view. The key frame extraction approach proposed in this paper is a two-stage process that key frames are first generated by the local camera and a decentralized process is followed to further reduce the number of key frames collected from a camera cluster in a decentralized fashion. To obtain local key frames, an online filtering mechanism with the Kalman filter is used to optimize key frames by searching the maximum of weighted importance of semantic features. Then we present a mechanism to select dominant camera and hand-off through the camera cluster so as to determine cluster key frame.
 
 

 
 
 

 
 

 

 

Conclusions 

    In this paper, we present a novel multi-view video summarization approach for decentralized key frame extraction in surveillance video contents for the purpose of dominant camera selection is presented. The result of experiment indicates that the features we adopt not only clearly represent the semantic of the object for key frame extraction, and it is also possible to couple camera selection methods together. Integration of dominant camera selection and key frame extraction presented employ an online video summarization system. This paper is useful to extend to more applications in surveillance system.
 
 
 
Demo Video
 
 
 
 
輔仁大學---智慧型系統實驗室