Research On Video Object Segmentation Algorithm Based On Spatio-Temporal Information Fusion

Video object segmentation is a binary label problem that separates the foreground object from the video background area.The difficulty of video object segmentation lies in the complexity of video content scene due to the richness of video content.At the same time,video often contains abundant motion information,which causes occlusion,fast motion and deformation between objects,and brings serious challenges to accurate and stable target segmentation tasks.In recent years,a great deal of video data has been accumulated,but manual processing has consumed a lot of manpower and financial resources.However,most of the existing research results still have serious limitations in terms of quality and efficiency.Fully utilizing the rich spatio-temporal information contained in video is conducive to solving the problems existing in video object segmentation.Therefore,the research of video object segmentation algorithm based on spatio-temporal information fusion has very important theoretical significance and practical value.In this thesis,several research perspectives of video object segmentation are studied,forming a research route from semi-supervised single object segmentation to unsupervised single object segmentation,and then to semi-supervised multi-object segmentation.The main research work of this thesis is summarized as follows.Firstly,an adaptive semi-supervised video object segmentation algorithm based on temporal continuity is proposed.Aiming at the inaccuracy of object segmentation in semi-supervised video object segmentation based on the measurement of object appearance and motion changes.This algorithm solves the difficulty that the network can not adapt to object changes well in semi-supervised video object segmentation,and improves the accuracy of segmentation.Secondly,in order to solve the time-consuming problem of CRF for post-optimization processing,a binary mask discrete point cluster optimization algorithm based on regional spatial correlation is proposed.As a lightweight algorithm,this algorithm reduces the time of post-optimization.The whole algorithm is compared with the current popular methods on DAVIS,GYGO,Segtrack-v2 and VideoSD datasets.The experimental results verify the effectiveness of the proposed algorithms.The proposed algorithm outperforms the popular methods in the accuracy of regional similarity segmentation and improves the accuracy to 0.883.Especially when the object in the video and the object in the given first frame have large outgoing or shape changes,the proposed algorithm has better robustness.Secondly,an unsupervised video salient object segmentation algorithm based on appearance and motion saliency is proposed.The algorithm integrates weak saliency detection module as an assistant to segment salient objects in video stream.Firstly,aiming at the insufficient utilization of video stream information in unsupervised video salient object segmentation,a novel self-growing video salient object segmentation algorithm based on appearance and motion saliency is proposed.This method is different from the currently segmentation method.In the way of network automatic growth,salient objects in video stream are gradually highlighted,and salient objects depend on the whole video stream.Then,aiming at the problem of region loss and whole incompleteness caused by pixel level segmentation,an optimization algorithm of region generation based on contour constraints is proposed.The algorithm is validated by experiments on DAVIS single object data set with the popular unsupervised video object segmentation algorithm.The segmentation accuracy of this algorithm reaches 0.706.,which is better than most of the contrast algorithms.At the same time,the proposed algorithm can improve the overall segmentation performance by replacing auxiliary modules with better performance.The proposed algorithm has more flexibility than the comparison algorithm.Finally,we study multi-object video segmentation algorithm based on semi-supervised single object segmentation network.We proposed a semi-supervised video multi-object segmentation algorithm.At the same time,in order to solve the problems of region overlap and occlusion in multi-object segmentation,this paper proposes a secondary classification based on appearance similarity measure,and the final classification of overlapping regions.Comparing the DAVIS video multi-object segmentation data set with the mainstream video multi-object segmentation algorithm,the validity of the proposed algorithm is verified in the experiment,and the segmentation accuracy reaches 0.643.This article consists of 25 drawings,9 tables and 70 references.