Research On Correlation Tracking Algorithm Via Attention Reinforced Semantical Feature

Although visual target tracking technology has made great progress in recent years,there is still a lack of high-performance tracker.Among them,the key issues that most of the trackers use shallow texture features instead of Reinforced semantic features,which play a key role in improving tracking performance.In order to solve the above problems,in this thesis,we develop correlation tracking algorithms via attention reinforced semantical feature,the main contribution of this thesis is summarized as follows:(1)Despite the demonstrated success of correlation filter tracking,However,the tracking performance is bad when the object is suffering from In-Plane-Rotation and Background Clusters in complex video scenes.Therefore,based on the attention mechanism,this paper proposes a hierarchical response fusion network.Firstly,the shallow texture features and deep semantic features are better obtained by the hierarchical convolution of the attention mechanism.Then,the feature is embedded in the correlation filter respectively,the template of deep semantic features learning makes up for the lack of discrimination ability of the shallow texture feature learning template.On this basis,the network adaptive search response map weight parameter makes the peak value of the response map maximized after fusion,and learns a more robust tracker.The experimental results show that the algorithm performs well on the OTB2013 and OTB2015 datasets,and the performance is greatly improved compared with the baseline algorithm.(2)Although the above method can get better results,it is not real-time,the drift of target is occurred when the object is suffering from occlusion and motion blur.For this reason,this paper further designs a network of end-to-end architecture is presented for deep correlation tracking via reinforced semantics and multi-attention learning.First of all,reinforced semantics obtains high-level semantic feature in the EDNet to make up for the drawback of individual low-level feature representation.Then,both channel-wise and spatial residual attention mechanism are leveraged simultaneously,which enables the network to extract more specific information for different tracking objects.At last,correlation filters layers are utilized to estimate the target locations according to the output maximum value in the response.Extensive evaluations are conducted on OTB2013 and OTB2015,the experimental results show that the algorithm is more robust and competitive than the most advanced tracker.