Unsupervised Video Object Segmentation Algorithm Based On Motion And Geometric Constraints

In recent years,along with the advance of the Internet and information technology,video and other data,as the main media of information transmission in the network,are growing by leaps and bounds,and have a profound impact on all fields of society.As one of the key technologies in video processing,video object segmentation plays a wide role in the fields of automatic driving,intelligent monitoring,3D reconstruction and video editing.Unsupervised video target segmentation task is to automatically segment the main foreground target in video sequence without a priori knowledge of segmentation target,which can save a lot of manpower and financial resources.It is one of the hot spots and difficulties in current research.This paper focuses on how to make full use of motion and geometry information to improve the performance of unsupervised video object segmentation.Firstly,a two-way optical flow motion optimization strategy combined with the previous frame propagation is studied to improve the segmentation accuracy in complex scenes.Then from the perspective of binocular study how to use the scene geometry information for unsupervised video target segmentation.Finally,the two algorithms are combined to study the unsupervised stereo video object segmentation algorithm based on motion guidance.The main work of this paper is as follows:(1)An unsupervised video object segmentation algorithm MGNet based on motion guidance is proposed.Due to the complexity of video scene,salient motion segmentation in a single frame is not always reliable.Based on this,this paper proposes a two-way optical flow motion optimization strategy combined with the previous frame propagation,through the network to learn the relationship between the forward and backward optical flow and the previous frame of the video sequence and the motion mode,so as to reduce the segmentation error.At the same time,a motion guided selective fusion strategy is proposed,which uses instance segmentation network to generate object suggestions and fusion the results of motion branches to remove the wrong motion background information and further improve the accuracy of segmentation results.The ablation experiments and comparative experiments on DAVIS-2016 and Seg Trackv2 dataset verify the effectiveness of the proposed algorithm mgnet.Among them,the average similarity of segmented regions on DAVIS-2016 is 77.5%,which is better than the mainstream similar methods such as UOVOS,MPNet,LVO and Fseg.(2)A novel unsupervised stereo video object segmentation algorithm Stereo Net based on disparity constraint is proposed.Aiming at the problem that the existing monocular video object segmentation network can not use the scene geometry information,an end-to-end unsupervised binocular video object segmentation network is proposed.The network extracts the scene geometry information from the binocular video image sequence,and fuses the parallax information of the corresponding frame with the multi-scale appearance features,So as to improve the classification accuracy of foreground target and background.In order to verify its effectiveness,the current mainstream public data sets of monocular video object segmentation are calculated and processed by using view synthesis technology,and the data sets DAVIS-2016-stereo and Video SD-stereo for binocular video object segmentation are constructed.The proposed algorithm,Stereo Net,is applied to the binocular video object segmentation data set.The average similarity of segmented regions on DAVIS-2016-stereo is 3.7% higher than that of the baseline network,and the average contour accuracy is 4.0% higher than that of the baseline network.(3)An unsupervised stereo video object segmentation algorithm Stereo MNet based on motion guidance is proposed.This paper introduces the results of the first research work into the unsupervised binocular video target segmentation network Stereo Net,designs the unsupervised binocular video target segmentation network Stereo MNet based on motion guidance,introduces motion information on the basis of disparity constraint,removes the wrong background segmentation,and further improves the accuracy of unsupervised binocular video target segmentation algorithm.The proposed algorithm stereomnet is tested on DAVIS-2016-stereo dataset,and the average similarity of the segmented regions reaches 75.5%.This article consists of 20 drawings,9 tables and 47 references.