Study On Human Instance Segmentation From Video With Fused Multi Clues

Foreground segmentation studies how the object is extracted from the video, it is afundamental problem in computer vision and is also the current popular research direction. In thevideo segmentation, the human instance segmentation has the special vital significance. Becausethe human body not only is representative in numerous non-rigid objects but also is focused on thecore in many video segmentation applications. For example object tracking in intelligent videosurveillance, pose estimation, motion analysis, human recognition, behavior analysis etc. All ofabove applications largely rely on video object segmentation technology. In human instancesegmentation from the video technology, in this paper, the research work mainly focused on thefollowing three aspects: video object segmentation technology based on the simple interaction,human instance segmentation with fused multi clues and implemented automatic human instancesegmentation system. The main research results include:(1) By the analyzing of existing video segmentation technique based on local color model, wedesign and implement foreground segmentation approach using locally competing1SVMs. Thecore of the method is to maintain two competing1SVMs models at every pixel, and the two1SVMs capture the local foreground and background color distribution separately. The usage oftwo local competing1SVMs can provide higher discriminative power and hand the ambiguousregion.(2)Propose a human segmentation approach with the fused multi clues. The approach useslocal color, object motion and shape information to segment video; finally it is solved by the graphcut framework. And these clues can be automatically weighted. Experiments show that our methodcan deal with background movement and complex scenes, and segmentation of the subsequentframe does not require user interaction.(3)Based on human body segmentation model using the fused multi clues and combined withhuman body detection technology, we design and implement an automatic human segmentationsystem. The system can initialize automatically and handle multiple human target segmentation atthe same time. The system is based on human body segmentation model using the fused multiclues, it uses HOG human detection method and Grabcut image segmentation method to segmentthe first frame. The system can also segment newly emerging human in the subsequent frames. Thepractice shows that our system is able to achieve good segmentation results for re-identification.