Research On Deep Learning-based Single Visual Object Tracking

In recent years,many deep learning-based tracking algorithms has shown excellent performance on various test platforms with the rapid development of deep learning technology and the disclosure of large-scale image data sets,but it is still difficult to design a robust tracking algorithm which is suitable for different application scenarios due to some reasons such as the variety of tracking objects and the appearance changes among sequences.For many discriminative methods which applied to RGB scenario,this paper designs a tracking network which combined the convolution network layer and the DAG-RNN layer to solve the problems of over dependence on features of the appearance and the weak discriminative capability of interference target.We use the undirected cyclic graph structure to approximate the image patch and decomposed it with directed acyclic graph,which used to model the dependency of the subblock's contextual neighborhood.This method enhanced the capability of inner-class discrimination and robustness.Besides,considering the influence of the motion state of target among sequences on the position of the predicted output bounding box,we utilize the motion revising layer to make the predicted bounding box more accurate.The experiment on OTB-50 and OTB-100 demonstrate that our method achieved an outstanding performance on Precision and Success aspects,especially in several challenging aspects such as BC and LR.In order to solve the problem of low real-time performance and excessive calculation loss which caused by utilized online fine-tuning strategy to obtain the general feature representation of different tracking objects by extracting a large number of positive and negative samples in discriminant method,this paper designs a real-time tracking network based on siamese architecture,which uses offline training and has a deeper network layer improved from existing methods.The weighted parameters are optimized in the pre-training stage by using end-to-end offline training.The results of EAO and A-R Rank evaluation and experiment on two assessment methods of VOT2017 show that our method can achieve real-time tracking performance with certain accuracy guaranteed.To deal with the problem that low robustness of feature extraction for the changing target which caused by the reason that generative methods based on template matching usually use the target feature of the first frame as the template,we improved existing methods and designed a siamese tracking method which can update the template in thermal scenario.The template can be updated by applying the template-updating strategy and background disturbance suppression method of optimal search area to the two branches of the siamese network,the target template can be update in real time by combining the previous frames and the disturbance suppression of the background area can be carried out according to the optimal target location.Besides,the TIR video sequences extracted from multiple open access thermal infrared image datasets are used to train the whole networks.A large number of comparative experiments and qualitative analysis on PTB-TIR and VOT-TIR show that our method can achieve superiority performance and advantages in different challenging aspects such as BC,MB and LR.