论文部分内容阅读
视频数据是信息存储的重要手段,但是很多时候(如机器人运动时采集的视频、监控相机记录的视频等),视频帧与帧之间相似度高,造成信息冗余,为信息的存储、查询、传输等带来困难.本文设计了一种语义浓缩算法,通过提取视频中的关键帧,实现快速、准确感知视频内容的目的.该算法首先采用背景建模方法从原始视频中提取包含重要信息的连续前景视频段,去除信息量较少的背景帧,然后通过采用视频语义分割方法将前景视频段分割得到一系列子视频,将从子视频的每一帧中提取出的特征构成原始字典,采用字典选择方法提取出关键帧.本文方法针对导航/监控视频,使用背景建模过滤无用信息,通过视频语义分割方法可处理不限长度的视频,而所选取的图像特征和字典选择方法则灵活有效地提取出有意义的关键帧.在实验中,通过对移动/固定平台视频的测试,并与人工标识及其他语义浓缩方法进行比较,证明本算法的重构误差较小,可以准确地提取视频中的关键帧.
Video data is an important means of information storage. However, in many cases (such as video collected during the movement of a robot and video recorded by a camera), the similarity between the video frame and the frame is high, resulting in redundant information for storing information and querying , Transmission and so on.This paper designs a semantic enrichment algorithm to extract the key frames in the video to achieve the purpose of fast and accurate perception of the video content.The algorithm firstly uses the background modeling method to extract the important information from the original video , The background foreground video segment is removed, and the foreground video segment is segmented to obtain a series of sub-videos by adopting the method of video semantic segmentation. The extracted features from each frame of the sub-video constitute the original dictionary, The method of dictionary selection is used to extract the key frames.This method filters the useless information for the navigation / surveillance video by using the background modeling, and can process the video of unlimited length through the video semantic segmentation method, while the selected image features and the dictionary selection method are flexible Effectively extract meaningful keyframe.In the experiment, through the mobile / fixed platform video test, and with Semantic station identification and other methods of concentration are compared, the algorithm prove small reconstruction error can be accurately extracted key frame video.