Candidate Region Proposal And Update Model Tracking Algorithm Based On Siamese

Visual object tracking is one of the important research directions in the field of computer vision.It is widely used in many fields,including video surveillance,automatic driving and military guidance.In the current object tracking methods based on siamese network,the quality of recommendation in candidate object regions is very important.The problem of model updating has also become an important research content in the corresponding object tracking methods.Most of the current tracking methods adopt anchor fixation as a recommended method.The number of candidate regions generated by this method is very large,but the quality is not very high.And in siamese networks,the update method of object template is usually linear update method.The tracker uses a simple linear update method to update the object model,but it can't deal with the object module with excessive appearance changes in the tracking process.Therefore,this paper mainly studies the object tracking algorithm of candidate region recommendation and template updating based on siamese network.The main work of this paper is as follows:1)This paper proposes a Heuristic Candidate Region Proposal for Siamese Visual Tracking Algorithm.The network model consists of two parts: template branch and detection branch.The template branch is mainly used to extract the features of the first or previous frame,while the detection branch is used to extract the features of the current frame.After extracting the features of the current frame,the probability graph method is used to predict the possible position of the object,and then the variable anchor method can be used to generate candidate regions around the position;then the main features of the corresponding region are recommended,and then through a feature adaptation network,the anchor points and the corresponding feature graph can further adapt to each other;finally,the template can be branched with the object The feature extracted by the detection branch is cross correlated,and then the maximum value of the response score is taken as the output.2)The template branch of SiamHCRP is modified to adaptive template branch to enhance the stability of network tracking.Based on SiamHCRP,modify the template branch and add an update component,Update Net.Update Net uses residual learning method,and then updates the corresponding training model.There are three templates in the whole network,which are the first frame template,the current video frame template and the cumulative template between them.In addition,the output of Update Net can be connected with the first frame template by using the jump connection method,and then the prediction of the next frame template is the output of the residual learning after the residual learning of the two templates.In order to verify the effectiveness of the algorithm,this paper mainly uses vot2016 data set and otb2015 data set for experiments.The experimental results show that the two algorithms are better than SiamRPN and SiamFC in tracking object.