时空对比度引导的视频显著目标提取模型
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

基金项目:

河北省重点研发计划项目(21351801D); 轨道交通关键装备智能运维平台研发(20310806D)


Video salient object extraction model guided by spatio-temporal contrast
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    针对经典视频显著目标提取模型没有充分利用时域显著性线索,易受背景噪声干扰,提取的显著目标不完整等问题,提出了一种在时空对比度指导下的视频显著目标提取模型。首先,自适应融合RGB颜色空间对比度和运动对比度,确定显著目标的先验信息;然后,利用当前帧的前景提取项和邻近帧位置约束项组成能量函数,指导时空显著性线索融合;最后,通过超像素级平滑优化处理,提取完整的视频显著目标。实验结果表明,模型在Visal,SegTrack V2和DAVIS数据集上进行测试,MAE值分别达到了0.030,0.024和0.032,F-measure值分别达到了0.772,0.781和0.812,具有良好的准确性和鲁棒性。因此,所提算法能够有效检测出视频中的显著目标,可为监控系统以及目标跟踪提供理论参考与方法依据。

    Abstract:

    The classical video salient object extraction model does not make full use of time-domain saliency cues,and is susceptible to the background noise interference.The extracted salient objects are incomplete.This paper proposed a video salient object extraction model under the guidance of spatio-temporal contrast.Firstly,adaptive fusion of RGB color space contrast and motion contrast was used to determine the prior information of prominent targets.Then,the energy function was composed of the foreground extraction item of the current frame and the position constraint item of the adjacent frames,which was used to guide the spatio-temporal saliency cue fusion.Finally,the complete video salient target was extracted by super-pixel smoothing optimization.The experimental results show that the model is tested on Visal,SegTrack V2 and DAVIS data sets.The MAE values in Visal,SegTrack V2 and DAVIS data sets are 0.030,0.024 and 0.032,respectively,and the F-measure values are 0.772,0.781 and 0.812,respectively,with good accuracy and robustness.This algorithm can effectively detect the visible targets in the video,thus providing theoretical reference and method basis for the monitoring system and target tracking.

    参考文献
    相似文献
    引证文献
引用本文

李春华,郝娜娜,刘玉坤.时空对比度引导的视频显著目标提取模型[J].河北科技大学学报,2022,43(2):144-153

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-10-12
  • 最后修改日期:2022-01-06
  • 录用日期:
  • 在线发布日期: 2022-05-03
  • 出版日期: