Research On The Algorithms Of Deep Neural Network Based Robust Visual Tracking

In this thesis,we first propose a robust deep long-term tracking(RDLT)algorithm.which can effectively alleviate the problems of long-term occlusion and out-of-view by incorporating an offline learned deep Siamese network into a unified correlation filter tracking framework.By employing both the correlation filter tracker and the Siamese network detector,we can obtain several target candidate windows.To correctly select the optimal target window,we further propose a novel multi-expert evaluation schema,which evaluates target candidate windows by a pair of "expert",i.e.,the correlation filter based expert(C-expert)and the deep similarity based expert(D-expert).Both the two experts jointly evaluate all the target candidate windows and select the optimal one as the final tracking result.Extensive experimental results show that the proposed RDLT algorithm performs favorably against the state-of-the-art tracking algorithms.Secondly,the deep features used in current correlation filter based tracking algo-rithms are high-dimensional and not well designed for the task of visual tracking.thus severely degrading the tracking accuracy and speed.To alleviate these problems.we propose a deep and shallow feature learning network(DSNet),which can effec-tively compress both low-level and high-level information of objects to uniform spa-tial resolution features,and learn the multi-level same-resolution compressed(MSC)features in an end-to-end manner.The learned MSC features can be incorporated into any correlation filter based tracking algorithm without any modification.In addition,we propose a channel reliability measurement(CRM)method to further refine the learned MSC features.Extensive experiments show that the learned MSC features have the advantage of allowing the equipped correlation filter based tracking algorithms to achieve better results while running at high frame rates.Thirdly,conventional classification based deep tracking algorithms are usually notdesigned for real-time applications due to the complex online sampling and optimiza-tion steps.To solve this problem,we propose a novel multi-stream auto-encoding generative adversarial network for efficient visual tracking.Different from the clas-sification based algorithms that treat visual tracking as an online learning and clas-sification problem,we propose a lightweight neural network based generator,which fuses multi-layer feature maps to accurately predict the target probability map in a single-shot generation.By applying the adversarial learning,the generator can learn more discriminative features.In online tracking,the learned generator can directly generate the target probability map corresponding to the searching region in a sin-gle shot.Experimental results show that the proposed algorithm can lead to the average tracking speed of 212 FPS on a single GPU,while still achieving favorable performance(57.2%DPR on UAV20L).Finally,we propose a novel hallucinated adversarial tracker.The proposed tracker mimics the human imaginary mechanism to generate diverse new positive training samples,which can effectively alleviate the over-fitting problem of tracking algo-rithms in online learning.Specifically,we propose a hallucinated adversarial network to learn non-linear deformations between a pair of same-identity instances.Then,a novel deformation reconstruction(DR)loss is presented to train the proposed hallucinated adversarial network in a self-supervised manner.To better select the deformations which are more suitable for transfer,we propose a selective deformation transfer(SDT)method.Finally,based on the hallucinated adversarial network,we present our hallucinated adversarial tracker(HAT).Experimental results on OTB-2013,OTB-2015 and VOT-2016 demonstrate the state-of-the-art performance of HAT.In particularly.HAT achieves the leading accuracy(95.1%)on OTB-2013.