Research On Improvement Of Target Tracking Algorithm Based On Transformer

With the advent of the era of big data,object tracking has attracted more and more attention in the field of computer vision,and object tracking technology has been widely used in people’s lives and industrial production.Target tracking research has achieved certain results and breakthroughs,but the complexity of the external environment and the influence of noise factors such as the deformation of the target itself make it still a challenging task to implement a robust target tracking algorithm.The challenge problems will help to improve the efficiency and quality of people’s lives and industrial production.One of the core issues of object tracking is how to effectively feature the object.The feature expression of the target refers to some useful information extracted from the appearance of the target,which is used to describe the shape,color,texture and other characteristics of the target.With the development of deep learning technology,researchers began to try to use deep learning models to extract features,using the powerful feature extraction capabilities of deep learning and the ability of transfer learning to build feature extraction modules for target tracking algorithms,so as to achieve better tracking.Effect.However,the current algorithm based on deep learning still has the problem of insufficient feature expression ability.To solve this problem,based on the Trans T algorithm,this paper improves the four parts of the algorithm’s feature extraction network and feature fusion network,and proposes an enhanced information Express Trans T tracking algorithm.The innovations and research contents of this paper are as follows:(1)In the feature extraction network of the Trans T algorithm,the Res Ne Xt network is used to replace the Res Net network,and the characteristics of the features in each group of feature spaces can be integrated through group convolution to improve the information expression of the model.(2)The NAM channel attention mechanism is added to the backbone network of the Trans T algorithm feature extraction network.The NAM channel attention mechanism can improve the performance of the model by enhancing the features with strong correlation and suppressing the features with weak correlation.(3)In the TransT algorithm feature extraction network,only deep features are used for feature fusion of the template area and the search area,while the shallow features extracted by the backbone network and the middle layer features are not fused,so part of the shallow layer information and the middle layer are lost.In response to this problem,the CEM contextual feature fusion network module is added after the Trans T backbone network to fuse shallow features,intermediate features and deep features,further improving the model information expression.(4)To solve the problem of insufficient enhancement of features by the MLP module of the feature fusion network part of the Trans T algorithm,the MLP module is replaced by the a MLP module with a single-head self-attention mechanism and a gating unit,which is adjusted by the gating unit.The impact on the input information,and establish the correlation of data elements in the feature sequence through the self-attention mechanism,which improves the effect of feature enhancement.The improved algorithm in this paper was evaluated on the OTB100 dataset.The experimental tracking success rate increased by 7.2%,the tracking accuracy increased by5.8%,and it can perform better in most challenging scenarios,which proves the improved Robustness and effectiveness of the algorithm.