Research On Video Object Segmentation And Tracking Algorithm Based On Graph Optimization And Mixed Graph Convolution Model

Object tracking has been active in the research of computer vision and has been widely used.In recent years,in order to reduce the impact of such problems as object occlusion,scale transformation,camera motion and object deformation on performance.The methods of tracking-by-segmentation have been proposed.However,there are still challenges such as non-rigidity and deformation in object segmentation and tracking tasks.Considering that the structured representation can further improve the robustness of the model while mining the potential data structure and context relationship of the graph.Therefore,graph based object segmentation and tracking has been widely concerned.In the task of video segmentation and tracking,video frames are generally divided into non-overlapping superpixels and composed as nodes,to achive object segmentation and tracking by the constraint and propagation of graphs.In order to further study the representation and learning of graph structured data,this thesis proposes a Graph mixed Convolutional Network representation model(Gm CN)and Graph optimized mixed Convolutional Network representation model(Gom CN),aiming to address the problem of object deformation and occlusion caused by non-rigid and fast motion.Two solutions can be listed as follows:(1)Considering that graph convolution can deal with irregular data,this thesis proposes a representation model based on mixed graph convolution(Gm CN).Firstly,the region of interest of each frame is divided into a certain number of superpixels by using common methods and extracts the visual feature representation information for each superpixel.Then,the smoothing and sharpening operation of graph convolution is used to realize more compact feature representation.Next,the optical flow information and context information are fused to ensure efficient use of the previous frame segmentation information.The spatial-temporal consistency constraint is explored to achieve more robust feature representation.Finally,this thesis adopts an effective optimization algorithm for Gm CN to obtain the global optimal solution.Experimental results of object segmentation and tracking on several standard datasets show that the tracking performance can be effectively improved by Gm CN.(2)Aiming at the problem of object segmentation and tracking in complex scenes,this thesis considers to use the optimized structure of graph to describe the potential relationships among superpixels.Traditional graphs(such as K-nearest neighbor graph,complete graph,etc.)lack of clear semantic structure.It is easy to be corrupted by redundant background information,which leads to the phenomenon of object drift.Structure optimization representation of graph is an important research task in the field of machine learning,It is widely used in various tasks.This inspires this thesis to further use the graph learning model combined with positive definite,low rank and nearest neighbor constraints to better learn the internal relationships among nodes.A better graph as the input of graph convolution will promote the aggregation of adjacent nodes.Based on this,a robust model Gom CN with good anti-noise ability is proposed to improve the performance of segmentation and tracking.Experimental results on several benchmark datasets show that the proposed model is also robust to the segmentation and tracking tasks with noisy background and non-rigid objects.