The Study Of Meta Learning Based Deep Neural Network For Visual Object Tracking

In recent years,as the basic technology of video and image sequence processing,visual object tracking has been gained extensive attention of researcher and become an important research area of computer vision.Visual object tracking has broad prospects and needs in many applications,such as city traffic control,intelligent navigation,visual interaction,intelligent robot,military field and so on.Therefore,the research of visual object tracking algorithm is of great significance to the construction of intelligent video image data analysis visual system.The task of visual object tracking requires to calculate the accurate location of a specified object in the image sequence or video with natural real scene's,and calculates the coordinates of the tracked object in the scene.Because in natural real scenes,the algorithm of visual object tracking needs to solve difficulties and challenges such as illumination change,tracked object occlusion,background clutter,violent motion of camera,similar interference objects,violent motion of tracked object,and severe deformations of targets.In thesis,we proposed a deep network based on deep convolutional neural network to directly calculating the object location coordinates of end to end,and will be applied with meta learning to makes the model can quickly learning tracked object representative features,to solve the object changes in the natural scene.The main contributions of thesis are as follows:1.Propose a novel network model for visual object tracking based on deep convolutional neural network,which can directly calculate the location coordinates of objects in scene pictures.The model employ the design of a prior box,which can judge the scale and the ratio of the width to height of the target,and make the final result of coordinate being more accurate.2.Apply meta learning to the proposed network model,to make the model can quickly learning the tracked object with differentiated feature expression,and it can adapt to all kinds of disturbances of natural scenes background.The network model training by meta learning method is better and more robust for visual object tracking.3.Design a complete tracking framework based on the deep network for visual object tracking.The tracking framework includes the modules of model parameters initialization,object location calculate,and model updating,and formed an end to end tracking algorithm.The algorithm can solve the target tracking problem of any video or image sequence,and the speed is real-time on a single graphic processor.