Research On Single Object Tracking Algorithm Based On Siamese Network

Single object tracking is one of the important branches of computer vision,and has always received special attention from scholars and experts.Its definition is to estimate the object position in subsequent frames given the object position information in the initial frame.Due to the influence of complex environmental factors on the single object tracking task,designing a tracking algorithm with both accuracy and speed is the difficulty of the current visual single target tracking task.Based on the siamese network structure,this paper conducts in-depth exploration and research on the difficult problems existing in single-object tracking tasks and proposes improvements,which improves the accuracy and robustness of the tracking algorithm and satisfies the real-time performance of the tracking algorithm.The specific research results are as follows :(1)A siamese network single-object tracking algorithm based on multi-scale representation and weighted fusion is proposed to solve the tracking drift problem caused by the insufficient extraction of object multi-scale features by the backbone network of existing tracking algorithms.The algorithm first uses a backbone network with multi-scale extraction capabilities as a feature extraction network,and improves the existing recurrent network paradigm,and proposes a Transformer weighted feature fusion recurrent network.The siamese network structure is improved from the backbone network and the fusion network level to meet the needs of tracking tasks in complex environments.The experimental results show that the algorithm shows good competitiveness in terms of accuracy and success rate.(2)A Transformer-based siamese network single-target tracking algorithm is proposed to solve the problem that most tracking algorithms ignore the feature extraction capability of the Transformer network.The algorithm first uses the Transformer network based on the selfattention mechanism as the backbone network,and taps the potential of the Transformer network in the field of tracking algorithms.In addition,since the tracking result obtained by predicting the image from the prediction network is relatively rough,the bounding box lifting module is used to perform fine-grained processing on the tracking result to achieve the refinement of the tracking result.Experimental results show that the algorithm has good tracking performance.(3)A siamese network single-object tracking algorithm for spatio-temporal template updating is proposed to solve the problem of adaptability to changes in target appearance caused by initializing the template algorithm for the first frame and not updating the template afterwards.The algorithm first uses the dynamic template update strategy to add dynamic template information every fixed frame to improve the robustness of the algorithm.Secondly,the dynamic scaling loss function is used to reconstruct the loss ratio of positive and negative samples,which improves the accuracy of the algorithm.Experimental results show that the algorithm balances precision and real-time performance,and has practical significance.