Object Segmentation And Tracking Based On Superpixel Graph Representation And Label Prediction Model

Object tracking is an important branch of computer vision and multimedia.The task of object tracking is to continuously locate and track one or more object targets in a video.It is widely used in road monitoring,human-computer interaction,remote sensing image analysis and many other fields.Although researchers have made great achievements in the field of object tracking,it is still a challenging topic.In the natural scene,there are a lot of factors that hinder the performance of the object tracking algorithm,including object deformation,illumination change,partial or complete occlusion,background clutter and so on.In recent years,object segmentation and tracking algorithms based on superpixels have been widely used in this field.There are some algorithms that mainly divide the image into many superpixels and construct graphs with superpixels as vertices and edges connecting vertices as edges,and realize object segmentation and tracking by the constraint and propagation of graphs.Due to the non-rigid and deformable objects in tracking,the constructed graphs may not be accurate.The algorithms based on graph learning are derived,which can better mine the potential manifold structure relationship of graphs.In this thesis,we focus on the problems of object deformation and background clutter,as follows:(1)For the object deformation problem in object tracking,we propose the label prediction model based on spatial-temporal consistency constraint.Firstly,the candidate region of interest(ROI)of each frame is divided into many superpixels which are used to extract feature descriptors and construct graphs.Then the score is calculated by support vector regression,and the weight of edge is calculated by Gaussian kernel function.The absorption time of each node is calculated according to the tracking framework of absorption Markov chain.Finally,by constraining the graph nodes with spatial-temporal consistency,we can get the optimal foreground probability and track the object after normalization.We have conducted a lot of experiments on the widely used benchmark datasets,and the experimental results show that the proposed model obtains good performance.(2)For the noisy background problem in object tracking,we propose a multi-view collaborative learning and label prediction model.Based on the above model,graph learning module and multi-view collaborative learning module are added to the model,and the graph structure relationship is updated iteratively in each frame.Firstly,the image is segmented into many superpixels which extract their features descriptors.Then the absorption time is obtained based on the absorption Markov chain framework.Finally,the optimized foreground probability is obtained through multi-view collaborative learning and label prediction to achieve object segmentation and tracking.We have conducted a lot of comparative experiments on several benchmark datasets,and the results show that the model significantly improves the performance of object segmentation and tracking.