基于超像素的视频多目标协同分割

Video object co-segmentation is an efficient method of video detection that automatically identifies common objects from multiple video clips. Video object co-segmentation research has a very broad application prospects in intelligent multi-video surveillance system. However, the existing video object co-segmentation algorithm lacks the theory of solving multi-object segmentation problem. Therefore, this project studies on multi-object co-segmentation algorithm based on superpixel: extend the video co-segmentation from pixel level to superpixel level, to improve the efficiency of computing; propose a superpixel feature based on illumination-invariant transformation, and the characteristics of the same common object from different video clips under different illumination conditions will be consistent; propose a co-saliency based on superpixel to describe the possibility of each region belonging to common objects, without the need to determine the number of common objects in advance; propose a multi-object occlusion processing based on optical flow vector, make full use of the correlation information between different video clips to deal with multi-object occlusion. The research will solve the problem of "inconsistency of illumination conditions between different video clips", "uncertainty of the number of common objects" and "diversity of multi-object occlusion problem" in video multi-object co-segmentation. The results will be applied to intelligent multi-video criminal investigation system to improve the accuracy and timeliness of multi-object tracking and detection.

视频目标协同分割是一种高效的视频检测手段，能够在多个视频中自动找出共同目标。视频目标协同分割研究在智能多视频监控系统中有非常广阔的应用前景。不过，现有的视频目标协同分割算法缺乏解决多目标分割问题的理论。因此，本课题研究基于超像素的视频多目标协同分割算法：研究将视频协同分割从像素级转化到超像素级，提高运算效率；研究基于亮度不变变换的超像素特征，使同一目标在不同光照条件的视频中特征保持一致；研究基于超像素的协同显著性，用于描述每个区域属于共同目标的可能性，无需事先确定共同目标的个数；研究基于光流矢量的多目标遮挡处理，充分利用视频间的关联信息处理多目标遮挡情况。本课题的研究将解决视频多目标协同分割中的“不同视频之间光照条件的不一致性”、“共同目标个数的不确定性”和“多目标遮挡问题的多样性”问题。成果将应用于智能多视频刑侦系统，提高多目标跟踪和检测的准确率和时效性。

视频目标协同分割是一种高效的视频检测手段，能够在多个视频中自动找出共同目标。视频目标协同分割研究在智能多视频监控系统中有非常广阔的应用前景。不过，现有的视频目标协同分割算法缺乏解决多目标分割问题的理论。因此，本课题研究基于超像素的视频多目标协同分割算法：研究将视频协同分割从像素级转化到超像素级，提高运算效率；研究基于亮度不变变换的超像素特征，使同一目标在不同光照条件的视频中特征保持一致；研究基于超像素的协同显著性，用于描述每个区域属于共同目标的可能性，无需事先确定共同目标的个数；研究基于光流矢量的多目标遮挡处理，充分利用视频间的关联信息处理多目标遮挡情况。本课题的研究将解决视频多目标协同分割中的“不同视频之间光照条件的不一致性”、“共同目标个数的不确定性”和“多目标遮挡问题的多样性”问题。成果将应用于智能多视频刑侦系统，提高多目标跟踪和检测的准确率和时效性。