Contrastive Siamese Collaborative Network For Visual Object Tracking And Application In Aircraft

Single object tracking as the hot spot direction,It usually builds the appearance and corresponding motion model of the object according to the target of the first frame of the video,and deduces the position of the object in the subsequent video.Many researchers have turned to convolutional neural networks,due to its impressive performance in visual tasks,to utilize its features for target tracking.A Siamese network-based tracking framework is crafted to utilize the features obtained by convolutional networks to differentiate the likeness between the sought-after image and the desired template.However,this way of comparing candidate frame and target at pixel level can not accurately track the target when dealing with complex scenes,this paper proposes an improved scheme of Siamese network object tracking based on contrast learning framework,which is applied to the aircraft tracking mission in order to address the issues of analogue interference,long-term occlusion,large-scale appearance changes,and other tracking challenges.The specific work contents are as follows:Firstly,different from the current Siamese network object tracking scheme,which generally adopts the pixel level target template matching,this paper proposes a Siamese network target template matching scheme based on contrast learning.Specifically,a special training set of target and analogue is constructed through the regional suggestion network,and the comparative learning and Siamese networks are trained cooperatively.In particular,a joint contrast Siamese loss function control model is used for training.In view of the collaboration of contrast learning,the proposed method can not only distinguish the target from the background,but also learn the essential characteristics of the target,so as to effectively distinguish the target from the similar in the complex scene.Experiments show that the proposed method achieves excellent performance,especially in dealing with the challenge of analogue interference.Compared with the basic method,the accuracy of OTB data set is improved by 23.79%.In addition,the challenges of long-term occlusion and large-scale appearance changes often occur in practical applications,which will lead to the discontinuity of the information changes in the semantic level of the object tracking algorithm.The tracking algorithm’s target template update is drastically impacted by this discontinuity,thus necessitating the implementation of a dynamic feature fusion object tracking scheme for aircraft tracking.This paper proposes such a scheme to be applied to the mission.Specifically,by adding a dynamic feature fusion module,the target template features and the target features of the search area are dynamically fused according to the weight factor of feature fusion based on the attention mechanism,so as to obtain the target template features containing different semantics and scales.In particular,in order to verify the effectiveness of the scheme,this paper constructs a dataset focusing on aircraft tracking.The experimental results show that the proposed scheme can well cope with the tracking challenges brought by occlusion and large-scale appearance changes on the open dataset,and achieve better performance on the constructed dataset.