| Visual object tracking has attracted much attention as an essential critical task in computer vision because of its fundamental theoretical value and comprehensive application demands.Due to a series of changes in the video sequence,such as occlusion,deformation,illumination change and so on,object tracking has been facing significant challenges for a long time.To overcome these challenges,with the rapid development of machine learning and artificial intelligence,many algorithms combining intelligent technology and object tracking have been proposed one after another.It promotes the development of object tracking that has increasingly become a research hotspot.The current mainstream tracking algorithm in deep learning adopts Siamese networks to identify and locate the target by matching template and search area.This method has a simple network structure and considers accuracy and speed simultaneously,achieving better tracking performance.However,due to the characteristics of Siamese network feature matching,it seriously depends on the representation ability of features.So,robust deep networks are needed to extract features.The tracking accuracy will also be reduced when the target is occluded and deformed,worst leading to severe tracking drift.In order to solve the above problems,this dissertation mainly focuses on the global semantic structure information and local sub-block location information of deep features and uses local information to make up for the deficiency of global information.At the same time,combined with graph network and attention mechanism to improve tracking accuracy and robustness.It mainly includes the following three parts:1.Semantic and context features integration object tracking algorithm is proposed.The object’s global features and local features are extracted and constructed respectively through the deep network.Both are organically aggregated so that the local features concretely describe the object,and the global features semantically and abstractly represent the object to construct an accurate and robust adaptive target representation.They feed back and supervise each other,realizing the tracking under the challenges of object occlusion,deformation,etc.;2.A center-suppressed graph convolution object tracking algorithm is proposed to fully mine the relationship between local features of sub-blocks.The graph network of sub-block features is studied to construct the graph network model of deep features and its adjacency matrix using cosine similarity and distance prior knowledge.Then,the deep local features are adaptively fused by the graph convolution operator.It can use the surrounding context information to make up for the missing information of object change,enhance the target representation and improve the tracking performance effectively;3.An attention mechanism object tracking algorithm is proposed,combining the object’s deep feature with the attention mechanism to simultaneously realize selfattention and cross-attention in the template and search area.Therefore,the sub-block feature,template and search area information can be fused with each other.The existing cross-correlation calculation method is replaced by the cross-attention,which can consider the correlation between feature nodes to further improve the network’s performance.We verify and compare our methods with the current tracking algorithms on the main public tracking datasets.Our proposed algorithms have reached the current advanced performance.For example,on the VOT2016 dataset,the expected average overlap(EAO)of semantic and context features integrated object tracking and centersuppressed graph convolution object tracking are 0.541 and 0.542,respectively,far more than other algorithms.The UAV dataset’s success rate and precision of the attention mechanism object tracking are 70.3% and 90.1%,respectively.Compared with other algorithms,it improves the performance by more than 2%.Extensive experiments show that algorithms we proposed have stable tracking performance and meet the real-time requirement,verifying the effectiveness of algorithms. |
