Research On Robust Visual Tracking Algorithms Based On Deep Learning

Visual object tracking is one of the most popular research topics in recent years,and it is still a difficult task to achieve fast,accurate and robust tracking due to challenges such as occlusion,illumination changes and fast motion.Deep learning-based visual tracking algorithms has attracted attention of researchers and has produced a large number of excellent algorithms,but there is still some potential for development.The main problems to be solved are: most of the features used for object tracking are single features extracted from the existing backbone network,without focusing on the important features that are beneficial to the tracking task,and it is difficult to give full play to the non-linear modeling ability of the network.Secondly,the appearance and size of the tracking object are constantly changing during the tracking process,and most of the object tracking algorithms use only the first frame as template,which cannot adapt to the changing object.Therefore,how to enhance the expression ability of the features and how to update the model are the problems that this article aims to solve.Aimed the above problems,this article improves the existing excellent tracking algorithm,in terms of feature enhancement and model update to improving the success rate and tracking accuracy of the algorithm,the main work is shown below.1.Aiming at the problem of object tracking failed due to incomplete feature utilization in complex background,a tracking algorithm based on higher-order statistics and template update is proposed.Firstly,the second-order information of the features is extracted to improve the similarity discrimination ability of the model.Secondly,a template memory update module is designed to determine the tracking status of the current frame by using the Io U average of multiple prediction boxes.Saved the frames with better tracking performance,and the templates are updated by merging the first frame and the saved,enabling the tracker to adapt to changes in the appearance of the tracked object,effectively improving the algorithm’s ability to adapt to dynamic targets.To evaluate the effectiveness of the algorithm,it is tested on OTB100,VOT2018,UAV123 and La SOT datasets,and the experimental results show some improvement in the tracking performance of the proposed algorithm.2.Aiming at the problems that the existing visual object tracking algorithms cannot fully utilize the local and global features of the object,a Transformer visual object tracking algorithm based on mixed attention is proposed.The algorithm makes the network more focused on the spatial and channel containing more object information by embedding mixed attention in the middle of network.Then,using dilated convolution to achieve multi-scale feature extraction of the image and enhance the local expression ability of feature.Finally,using Transformer’s global modelling ability to transmit feature information between the two branches of the Siamese network by Encoder-Decoder.In Transformer,the large convolution kernel is used to giving each feature a more flexible and efficient position encoding.The proposed algorithm has been tested on the OTB100,VOT2018,and La SOT datasets,and the results show that by exploiting the Transformer architecture,the accuracy and success rate of the algorithm has improved,especially on the long-time tracking dataset.3.Aiming at the problem that the object tracking algorithm does not specifically deal with the feature of each layer of the backbone and the independent optimization of the two branches of classification and regression,a hierarchical feature enhancement object tracking algorithm based on the anchor-free is proposed.The method uses a specific feature refinement module to process and fuse the features of different layers from the backbone,giving full play to the advantages of each layer’s features,and improving the network’s feature expression ability and tracking accuracy;Secondly,the connection between classification branch and regression branch is established to eliminate the problem of unbalanced training between regression and classification branch.The experimental results show that as a short-term tracking algorithm,it has certain competitiveness in the accuracy,success rate and speed on short-term datasets.