Research On Siamese Network Object Tracking Based On Semantic Fusion And Template Update

Visual object tracking technology is one of the basic research directions of computer vision,which has been widely used in many fields.Object tracking can be regarded as a task of continuously locating the position,scale and trajectory of the object in subsequent video sequences based on known object characteristics.In recent years,object tracking algorithms based on the siamese network series have been proposed continuously.The tracking task is divided into two stages:offline training and online tracking.The tracking performance has surpassed the traditional correlation filtering object tracking algorithm.The trackers based on siamese network have achieved excellent tracking performance at present.However,when the algorithm encounters the challenges of illumination change,occlusion,excessive deformation and scale change,there are still problems such as poor prediction box and tracking drift.In order to solve the above problems,this paper conducts research from two aspects:algorithm integration and template update.The specific work is as follows:(1)This paper proposes a Semantics-SiamRPN algorithm for tracking and correction using target semantic information.The algorithm first calculates the semantic information of the target by detecting the network.On the basis of semantic information,the similarity prediction method of two-stage target optimization is used to eliminate the interference target.According to the similarity matching method,the potential target and the three-color feature of the historical tracking result are matched and calculated,so as to relocate the target position and correct the deviation of the tracking results.In addition,in order to improve the semantic recognition ability of the detection network,this paper conducts transfer learning training on the detection network.This paper conducts experiments on the OTB 50,OTB100,and VOT2018 datasets.The results show that the Semantics-SiamRPN algorithm proposed in this paper is significantly better than the baseline algorithm DaSiamRPN and other international cutting-edge algorithms.(2)The prediction results of the siamese network tracking algorithm in complex scenes such as occlusion,short-term violent deformation,and background interference are easy to introduce irrelevant background information.The UpdateNet algorithm updates the prediction results of each frame to the template,which affects the purity of the tracking template.In order to solve the problem,this paper proposes a template sparse update method based on reliability evaluation.In this method,the reliability evaluation mechanism is first used to evaluate the confidence of the predicted results,and then sparse update conditions are set according to the evaluation results to abandon the update of the tracking results with low confidence.Finally,this paper conducts experiments on the VOT2018 dataset,and the results show that this sparse update method ensures the purity of the tracking template to a certain extent and improves the performance of the tracking algorithm.