On-line video multi-object segmentation based on skeleton
model and occlusion detection
Received: 11 October 2017 / Revised: 30 April 2018 / Accepted: 23 May 2018
Springer Science+Business Media, LLC, part of Springer Nature 2018
Abstract In this work, we propose an approach for on-line video multi object segmentation
based on skeleton model and occlusion detection. We consider the multi-object segmentation
in every frame as a multi-class region merging based object segmentation. We then generate
the initial object superpixels automatically using a skeleton model from the second frame.
Moreover, we also propose an initial background superpixel prediction scheme. In case the
occlusion to affect the final segmentation result, we propose an occlusion detection model
based on optical flow. The experimental results show that our method is both robust in
segmenting multi objects and efficient in execution time.
Keywords Multi-object segmentation
The target of video object segmentation (VOS) is to extract the primary object from
video clips and to eliminate the background in all frames. Video object segmentation is
one of the hottest research topics in video processing, and it has many applications in
surveillance, video editing and video matting etc. Video object segmentation has re-
ceived a considerable attention from the video processing community in recent years and
various methods have been proposed [3, 5, 9, 16, 18, 21–24, 27]. For instance, Lee et al.
 present a method which first identify object-like regions according to both static and
Multimed Tools Appl
* Chi-Man Pun
Guangdong University of Technology, Guangzhou, China
University of Macau, Macau, SAR, China