Research On Algorithms Of Automatic Video Object Segmentation In Complex Scenes

With the continuous upgrading of infrastructure of Internet and the rapid spread of mobile terminal equipment,it is more and more convenient for people to shoot and watch the video.Because of the contained information is rich and vivid,video has be-come one of the most important information carrier in life.Increasing massive video data also brings the demand to identify,retrieve and understand video content.How to reduce the difficulty of video content to understand,extract the key information in the video has become the important research topic of video processing.As it aims at effec-tively segmenting the salient foreground object,video object segmentation has a wide range of applications on video summation,video retrieval,motion analysis and video semantic understanding.The current video segmentation algorithms are mostly bottom-up approaches,which segment the salient foreground object by acquiring and analyzing the bottom features such as color,edge and motion information in the video.The traditional algorithms based on artificial annotation can not meet the application requirements of the current mas-sive video data environment.As the scene and shooting condition of massive video are complex and diverse,the current algorithms of automatic video object segmentation can not maintain robustness in some complex scenes.In view of the above problems,this paper proposes two algorithms of automatic video object segmentation for differ-ent scenes.The main work and innovation of this paper is as follows:1.The existing algorithms based on Graph Cut is easy to be disturbed by back-ground noise and pixel mismatch,and the robustness of them is poor in some complex scenes.In this paper,an automatic video object segmentation algorithm based on optical flow field and Graph Cut is proposed to improve the above problems.Before the foreground object segmentation,the proposed approach an-alyzes the global motion features of the video,obtains the prior knowledge of the foreground object,and reduces the interference of the background noisc,Aiming at resolving the problem of pixel mismatch,the proposed approach proposes a dynamic position model optimization,and leverage the position model of the foreground object to enhance the temporal consistency of the segmentation re-sult.Experiments show that the proposed approach can obtain more accurate and robust segmentation result in the scenes that the shape and motion of foreground object is irregular.2.In some complex scenes,the existing algorithms based on object proposal tends to be partially missing.The reason of the problem lies in the fact that the object proposal is too fragmented and the mapping of time domain between object pro-posal is not accurate.This paper presents an improved algorithm based on object proposal.The proposed approach not only improves the fragmentation of object proposal,but also improves the temporal consistency of object proposal between adjacent frames.In order to further enhance the accuracy of the mapping of time domain,the proposed approach introduces more features to measure the weight of edges.Experiments on multiple benchmark data sets show that anti-noise ability of the proposed approach to background noise is stronger than existing algorithms,and the segmentation result is more complete in complex background environ-ment and surface reflection.