Research On Siamese Network Tracking Algorithm Based On Graph Attention And Residual Learning

Visual object tracking is an important and challenging task in computer vision,widely used in autonomous driving,video surveillance,and robotics.Siamese trackers have received increasing attention in visual tracking due to their high accuracy and tracking efficiency.However,most of the correlation operations of Siamese trackers directly use template features to slide windows on search region features,without distinguishing the discriminative parts of objects from background noise,blurring the spatial information of response features.In addition,the bias between classification confidence and regression accuracy that exists in classification-regression networks after similarity matching also limits the tracking accuracy of trackers.This paper mainly studies the improvement of the tracking algorithm based on the twin network.The specific research contents are as follows:(1)Aiming at the problem that the Siamese tracker uses related operations for similarity matching,resulting in fuzzy spatial information of response features,a target semantic guided graph attention feature fusion network is proposed,which effectively removes the fixed-size templates used by related operations by using adaptive templates.Background noise,the network also models the contextual semantic relations of objects and uses the resulting semantic relations to guide the feature fusion process in a part-based manner,thus accurately highlighting discriminative parts of objects.It can not only remove background information,but also focus on the discriminative part of the target object,thus effectively solving the problem of fuzzy response features caused by correlation operations.Finally,an object-aware prediction network is used to learn object-aware features for classification and regression tasks,which effectively improves the discriminative ability in tracking.(2)Aiming at the serious deviation between classification confidence and regression accuracy caused by independent optimization of classification and regression in Siamese networks,a Siamese network tracking method based on reciprocal positioning-aware classification and residual regression is proposed.This method constructs a residual regression network,refines the bounding box through a residual learning mechanism,and fully exploits the accuracy of object localization.And a localization-aware classification network is designed to use the predicted regression results to guide the classification confidence and improve the foreground-background classification ability.The interaction of residual regression network and localization-aware classification network forms a closed-loop structure to fully exploit the mutual benefits between classification and regression tasks.Therefore,the consistency between classification confidence and regression accuracy is guaranteed during tracking,leading to higher tracking accuracy and execution efficiency.Finally,in order to verify the effectiveness of the proposed method,comparative experiments and ablation experiments are carried out with the recent Siamese tracking method on the challenging tracking dataset.The experiments show that the proposed method can avoid the interference of background noise and accurately locate the target.