Object Tracking Based On Semi-Nonnegative Coding And Spatial Constraints

Object tracking,as an important research direction in the field of computer vision,has always been the concern of domestic and foreign scholars.The main task of object tracking is to locate and track the target in the video sequences.Object tracking has a wide range of practical applications,such as video surveillance,autopilot,intelligent robots,video communications and so on.Due to the complex and changeable application of the object tracking,most of the tracking algorithms are only applicable to specific scenes.The research of the tracking algorithm still needs to be further studied.A well-designed tracking algorithm requires both robustness,real-time and accuracy.In this thesis,a series of challenging problems such as occlusion,deformation,complex background and so on are analyzed,and the main research results are as follows:1.An object tracking method based on Semi-NMF integrated learning is proposed.This method trains a number of weak classifiers,and obtains the global optimal subset of weak classifiers by Semi-NMF coding.These weak classifiers are combined into a strong classifier by linear weight,which can locate the object.The classifier parameters are updated online by the predicted object location to ensure that the object appearance model is adapted to achieve accurate object tracking.2.A Semi-NMF coding object tracking method based on NL-means is proposed.This method focuses on the construction of interframe constraints based on NL-means.The sensitivity of the tracker to the small changes and partial occlusion of the adjacent frame is weakened by the similarity of the features within the inter-frame superpixel neighborhood.Moreover,the stability of the coding model is enhanced by Semi-NMF coding.The continuous updating of the codebook and the classifier parameters ensures the robustness of the tracking,so that the object can be tracked accurately when faced with changes in illumination,occlusion,and so on.3.An object tracking method based on spatial constraint support vector is proposed.This method mainly improves the accuracy of the tracking algorithm by optimizing the training sample sets and the support vector sets by spatial constraint.Firstly,we construct the object confidence map,and give weights of the sample features according to the spatial position.At the same time,the weights of the support vectors are allocated through the space distance constraints to maintain the tracker performance.