一种用于视频浓缩的图像匹配算法

An Image Matching Algorithm for Video Synopsis

为了保证公共安全,身边的摄像头越来越多,在录像监控视频中提取有效信息也变得日益困难。针对该问题,提出一种新的基于无监督学习的图像匹配算法。该算法无需像以往算法一样,需要大量人工标注的数据,而是通过观看视频即可完成训练,从而节约了大量时间成本。利用图像匹配算法,可以将视频中无效与相似内容进行浓缩,以提炼出有效内容。该算法在MPI-Sintel与KITTI集上进行了测试,APE分别为4.695与5.838,Accuracy@30分别为0.993与0.967,达到了较为理想的效果。

To ensure public safety,there are more and more cameras around us.Along with that,there are more and more long-term video surveillance.It is increasingly difficult to extract effective information in these videos.To solve this problem,this paper uses a new image matching algorithm based on unsupervised learning.It does not need the same amounts of manually labeled data as the pre-vious algorithms.It can complete training by simply watching the video and save a lot of time and cost.Then through the image matching algorithm,invalid and similar content in the video can be condensed to extract effective content.Through experiments,the algorithm of this paper was tested in the MPI-Sintel and KITTI sets,the APE are4.695and5.838and the Accuracy@30are0.993and0.967.Therefore good results are achieved.