基于视觉显著性车辆监控视频关键帧提取方法-《计算机技术与发展》

文章信息/Info

Title:: A Key Frame Extraction Method of Vehicle Surveillance Video Based on Visual Saliency

Author(s):: ZHONG Meng-jie; ZHANG Yan-bin; School of Telecommunications &Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China

Keywords:: vehicle monitoring video; key frame extraction; low-level feature; license plate similarity map; weighted average combination

摘要:: 针对道路车辆监控视频中车辆的关键帧提取问题,基于视觉显著性提出了一种关键帧提取方法。该方法首先对监控视频中的目标车辆进行检测、跟踪及背景分离,其次分别提取目标车辆的 FT、LBP 与边缘三种底层特征,并在 RGB 颜色空间根据改进后的加权欧氏距离得到车牌相似图。然后根据车牌相似度越小则给三种底层特征越大惩罚的思想,采用车牌相似图对 FT、LBP 与边缘三种底层特征图像进行优化,将三种优化后的底层特征图像进行加权平均融合,得到多特征融合图像。最后以多特征融合图像结果为参考,选取出目标显著性程度最大的融合图像对应的视频帧为关键帧。实验结果表明,该方法能提取监控区域的序列图像帧中车辆信息最丰富的图像,实现道路车辆监控视频中车辆数据的有效压缩。

Abstract:: A key frame extraction approach based on visual saliency is proposed aiming at extraction of the key frame of vehicle on road surveillance video. Firstly,the target vehicle in the surveillance video is detected,tracked and separated,then the FT,LBP and edge features of the target vehicle are extracted and the similarity map of the license plate is obtained according to the improved weighted Euclidean distance in RGB color space. Secondly,according to the idea that the smaller the similarity of license plates,the greater the penalty for the three underlying features,three underlying feature maps are optimized according to the similarity map of the license plate,and the multi-feature fusion image is obtained by weighted average fusion of three major feature maps. Finally,taking the best definition of the multi-feature fusion image as the reference,the frames whose fusion maps can pop out the objects best are selected as key frames. Experiment shows that for a series of vehicle images in the surveillance video,the algorithm presented effectively selects the most abundant image of vehicle,realizing effective compression of vehicle data for road vehicle surveillance.