Deep Network-based Researches For Robust Visual Tracking

Visual tracking is considered as an important research orientation in computer vision,it has the profound research significance and the far-reaching application prospect in video surveillance,intelligent transportation,visual navigation and military guidance.The essence of visual tracking is to predict the state of specific target in the consecutive video sequences.However,it may face many difficulties and challenges to achieve the tracking methods with high accuracy and strong robustness in the actual tracking environment.This paper lucubrates the tracking algorithms from the perspective of network construction and appearance characteristic modeling to effectively alleviate target drift in complex tracking scenes.To address the problem that ambiguous samples degrade the reliability of network during the long-term tracking,this paper proposes a novel tracking method based on the reliability verification network.The reliability verification network can connect with the arbitrary convolutional neural networks by sharing convolutional layers.The dual decision mechanism can measure the reliability of current forecasting result and then correct deviation to avert excessively accumulating error and degrading discriminative ability.To make the reliability verification network estimate the similarity between prediction and real target more accurately,the feature selection model and similarity fusion strategy are adopted to optimize the appearance characteristics and similarity labels respectively.The experimental result shows that the verification network can improve the accuracy and robustness to effectively alleviate drift.To address the problem that the appearance characteristic model of region proposal network has the weak ability to discriminate the foreground and semantic background,the paper proposes an attention-based multi-scale tracking algorithm.The attention mechanism mainly includes spatial and channel attention networks.Specifically,the spatial attention network learns the planar weights to enhance foreground and suppress background information,and the channel attention network learns the dimension weights to discard redundant noisy feature maps.Besides,the spatial and channel attention networks focus on appearance similarity and semantic classification characteristics respectively.Experimental results show that this method can learn the otherness between foreground and background to improve discrimination and alleviate drift.