| Visual tracking is one of the most fundamental problems in the field of computer vision,and it has a wide range of applications such as autonomous driving,UAV applications and military reconnaissance.In practical use,there are many challenges,such as target occlusion,scale variation,and untimely target disappearance and reappearance.Currently,short-term tracking has achieved very good tracking performance,but when the target is completely lost or out of view for a long time,short-term tracking will not be able to handle these situations,making long-term tracking more valuable for practical applications.In order to achieve stable long-term tracking,the ability of the visual tracking algorithm to accurately determine whether the target is lost is one of the keys.Once it determines that the target is lost,the tracker will perform re-detection and tracking in a timely manner.There are many determination methods available,but a common problem with these methods is that they require artificially given determination thresholds,and it is a challenge to choose the appropriate thresholds.To address the above problems,we focuses on proposing three deep learning-based target drift discriminative networks to determine whether the tracking target is lost in a timely manner without artificially selecting a threshold value,and establishing three longterm tracking algorithms.The main contributions of this thesis are summarized as follows:(1)Aiming at the problem of misjudgment in visual tracking due to high target response values,but not tracking results.We design a target drift discriminative network without a threshold to determine whether the target is lost or not.The network uses four convolutional layers,three full connection layers and Softmax function to judge the tracking results.When training the network,established positive and negative samples are used to obtain better target discrimination.Finally,a long-term tracking algorithm is constructed by introducing a target drift discriminative network into the base tracker.When the target is judged to be lost,a new search area will be selected to find the target quickly.The method was tested on datasets such as UAV20 L and VOT2018-LT.Compared with several other classical threshold discriminative criteria,we do not need to set the threshold artificially and has better judgment performance.(2)Aiming at the problem that the targets are repeatedly lost,reappearance and have huge deformation in long-term visual tracking.A target drift discriminative network based on a dual-template Siamese structure is designed.The network uses both static and dynamic templates to jointly determine whether the tracking results are lost or not,and without the need for artificially determined thresholds.Among them,the introduction of dynamic templates effectively improves the algorithm’s ability to adapt to changes in target appearance.In order to train the proposed target drift discriminative network,a sample-rich dataset is established.To verify the effectiveness of the proposed network,a complete longterm tracking algorithm is constructed by combining this network with the base tracker and the re-detection module.It is also tested on classic visual tracking datasets such as UAV20 L,La SOT,VOT2018-LT and VOT2020-LT.The experiments show that our method has significant advantages in long-term tracking.(3)In order to make full use of the feature information between different templates to improve the discriminative power of the target drift discriminative network.A target drift discriminative network based on feature fusion is designed.The network uses static and dynamic dual templates as reference templates and performs channel-by-channel multiplication with the initial tracking results respectively to retain salient feature information and suppress useless information.Since there are factors such as deformation and occlusion of the target,we then introduce a refinement module to enable accurate localization of the target position within a local range.Finally,in order to track the target quickly and accurately,we propose a switching strategy between local and global tracking.Experiments show that our method has the best performance on classical visual tracking datasets such as UAV20 L,VOT2018-LT and VOT2020-LT. |
