Visual Tracking Algorithms Based On Convolutional Neural Networks

Object tracking is one of the most popular task in computer vision and is of great importance to areas such as automatic driving,intelligent surveillance and et al.Although many algorithms have been proposed,there are still challenges remaining in this area.With the development of deep learning,improving tracking performance by applying neural networks has become a hot topic.According to this,contributions of this thesis are shown as follows:(1)A correlation filters based long-term tracking algorithm is proposed.Firstly,convolutional features from different layers are ultilized to improve tranditional correlation filters' power.Secondly,features from the context are integrated into the filters to deal with boundary effects.Thirdly,a SVM based re-detector is adopted to cope with heavy occlusion and target disappearance.(2)An adversarial learning based fully-convolutional Siamese networks for tracking is proposed.The generative networks learn to produce masks which are used to augment training samples.With augmented samples,the feature extraction networks is supposed to avoid overfitting and achieves more powerful representations.(3)A densely connected networks is proposed to improve the tracking performance of Siamese Region Proposal Networks.A deeper networks is built to extract more powerful features.As the deep semantic features are too sparse to predict the position,direct connections are introduced from any layer to all subsequent layers.In order to prevent zero paddings from making displacements,a crop layer is introduced to remove the outliers of feature maps.Then,the Region Proposal Networks is used to make a more precious bounding-box regression.(4)Experimental results on OTB-2013 and OTB-2015 reveals that the proposed trackers perform well on accuracy and robustness.