Visual Tracking Based On Deep Learning

Visual tracking is an important research topic in computer vision with numerous ap-plications including surveillance,robotics,human-computer interface and behavior analysis.Due to several complication factors,such as background clutter,illumination variation,par-tial occlusions and deformation,robust tracking is still a challenging task after decades of research.Benefiting from the emergence of large-scale visual data such as Image Net and the continuous improvement of computer performance,the application of deep neural networks,especially CNNs,have been developed and achieved state-of-the-art performance in many vi-sion task,such as image classification and object detection.It also offered new opportunities for the research of tracking.Under the framework of deep learning,this dissertation tries to explore the application of several different deep neural networks in the visual tracking by taking advantage of the powerful characterization of deep neural networks.The main contributions of this disserta-tion include the following aspects.We use pre-trained neural networks in combination with traditional tracking method.We present an adaptive tracker which integrates the kernel correlation filters with multiple effective CKN descriptors.By adopting a Flip Flop scheme,the weights of different features can be adjusted in the process of tracking to get better performance.Extensive experimental results on the OTB2013 tracking benchmark show that our approach performs favorably against some representative state-of-the-art tracking algorithms.We train a siamese network offline for online visual tracking.In this dissertation,a tracking algorithm based on siamese network is proposed.Aiming at the existing problems of siamese network based tracking methods,a normalized convolution operation is proposed.Through the normalization of L₂in each receptive field in the convolution operation,the network can obtain stronger discriminating ability through offline training.In addition,by expanding and rearranging the data through the training set,the problem of uneven distri-bution of training samples is solved to a certain extent.In order to cope with the change of appearance of targets in the tracking process,a heuristic template updating method is introduced,which updates the network weights of offline training online by dynamically up-dating the tracking training set.The results on multiple benchmarks demonstrate the benefits of normalized convolution and online updates.We train and fine-tune a fully convolutional network online for robust tracking.We propose a Part-based convolution Network(Part Net)framework for tracking,which incor-porates both holistic and local information of all candidates.Firstly,we use the pre-trained CNN as the basic network to extract the deep features.Secondly,a set of locally shared convolution kernels is introduced to incorporate local features with the global one for a more discriminatory feature map.Thirdly,the offset of each sample and parts are used to train the network,which combines both high-level semantic context and low-level spatial structure.Lastly,our Part Net is updated online with a periodic schemes to address drift and occlusion problems in tracking scenario.