Research Of Visual Object Tracking Method Based On Semi-Supervised Learning

Object tracking is a hot issue in the field of computer vision and plays an important role in intelligent traffic,human-computer interaction,security monitoring and so on.The task of object tracking is to locate the object in each frame of an image sequence,and provide a complete object area to generate its motion trajectory.However,due to the presence of occlusion,indistinct background distinction and illumination changes in the video,object tracking is still a difficult task after decades of research and development.In the absence of sufficient training samples,the performance of the traditional object tracking method based on supervised learning will be greatly affected.Semi-supervised learning is one of the hot topics in the field of machine learning.For many learning tasks,collecting tagging data involves expertise,making it costly.Collecting unlabeled data is easier and cheaper.Through semi-supervised learning,unlabeled data can be fully utilized,which greatly reduces the demand for labeled data.By combining semi-supervised learning with object tracking,the unmarked samples obtained in the tracking process can be fully utilized to optimize the model and improve the accuracy in the case of limited annotated data.Aiming at the single object tracking problem in complex environment,this thesis studies the object tracking method based on semi-supervised learning,and proposes and implements two object tracking methods based on semi-supervised learning,namely the object tracking method based on graph-regularized fuzzy least squares support vector machine and the object tracking method based on multi-view collaborative learning.(1)In object tracking method based on graph-regularized fuzzy least squares support vector machine method,use map separately from regular mining structure information,time and space to build video sequence of time and space structure,and space-time structure information as the constraint conditions of object tracking,combining fuzzy learning,building common graph-regularized fuzzy least squares support vector machine,and from two angles of dual form and representation theorem to solve the objective function,to realize object tracking algorithm.(2)Based on the multi-view collaborative learning strategies,with the related filter framework,we uses multiple layers of deep convolution neural network to construct filter sub models,and puts forward two kinds of adaptive weighted strategy,adaptive allocation weights for each model,the multiple-view submodel weighted combination,and uses the normalized peak value to estimate the tracking reliability,realize the object tracking model of joint,said to build a more accurate apparent purpose of the model.The performance of the object tracking algorithm can be improved effectively by the collaborative learning decision.The two object tracking algorithms based on semi-supervised learning proposed in this thesis have been tested on OTB100 dataset.The experimental results show that the object tracking methods proposed in this thesis have good performance in accuracy and success rate.