Research On Siamese Networks Based Object Tracking Algorithm Research

Visual object tracking has received increasing attention over the last decades and and is widely used in many visual applications such as surveillance,robotics and human-computer interaction.Although much progress has been made in related research,it is still a key issue to reduce the effects of lighting changes,occlusions,and other factors,and to ensure the real-time performance of target tracking.In recent years,convolutional neural networks demonstrated their superior capabilities in various vision tasks.They have also significantly advanced the state-ofthe-art of object tracking.Some trackers integrate deep features into conventional tracking approaches and benefit from the expressive power of CNN features.Some others directly use CNNs as classifiers and take full advantage of end-to-end training.Most of these approaches adopt online training to boost the tracking performance.However,due to the high volume of CNN features and the complexity of deep neural networks,it is computationally expensive to perform online training.As a result,most online CNN-based trackers have a far less operational speed than real-time.In response,siamese trackers came into being,and have received much attention due to their high speed and accuracy.These Siamese trackers formulate the visual object tracking problem as learning a general similarity map by cross-correlation between the feature repre-sentations learned for the target template and the search region.However,when the target has obvious appearance changes or has a cluttered background,the generalization and discriminability of the tracker is still poor.To improve the generalization and discriminability of the tracker,this paper provides two structures for target tracking based on siamese networks.Because in a deep CNN trained for image classification task,features from deeper layers contain stronger semantic information and is more invariant to object appearance changes,but being discriminative allows the tracker to differentiate the true target from the cluttered or even deceptive background is still poor.In this regard,this paper proposes an object tracking algorithm based on attention-guided fully-convolutional siamese networks,which aims to improve the discriminability of the tracker(SiamFC).At the same time,we noticed that the backbone networks used in Siamese trackers are relatively shallow,such as AlexNet,which does not fully take advantage of the capability of modern deep neural networks.And direct replacement of backbones with existing powerful architectures,such as ResNet and Inception,does not bring improvements.Therefore,proposing a residual siamese network for real-time object tracking,which uses deeper convolutional neural networks to enhance tracking accuracy and generalized,and achieve real-time object tracking.details as follows:(1)Object tracking algorithm based on attention-guided fully-convolutional siamese networks: Aiming at the problem that the tracking effect is not good when there is a complex background,that is,there are too many similar objects.we design a channel attention module.And embed it into one of the branches of the tracking network(SiamFC),channel-wise weights are computed according to the channel activations around the target position.The inherited architecture from SiamFC allows our tracker to operate beyond real-time.The Got-10 k data set is used in conjunction with the newly proposed evaluation method to verify the proposed object tracking algorithm,and its performance is better than tracker SiamFC.(2)A residual siamese network for real-time object tracking: However,Siamese trackers still have an accuracy gap compared with state-of-the-art algorithms,and they cannot take advantage of features from deep networks to improve the tracking accuracy.To address the issue,adopt new residual modules and further design a new residual network architecture using these modules,We replaced the backbone network in the siamese network tracker CFnet with the network architecture CIResNet-22.The designed architectures are lightweight and guarantee real-time tracking speed.Finally,on the data set OTB-2013,Two metrics,i.e.precision and area under curve(AUC)of success plots,are used to rank the trackers.Experiments proved the proposed method is better than other methods.