Siamese Network Based Real-Time Visual Object Tracking Algorithm

Visual object tracking is one of the popular research topics in computer vision.It has a wide range of application scenarios in the fields of intelligent monitoring,human-computer interaction,and security monitoring.In recent years,although researchers have proposed a large number of excellent visual object tracking algorithms,developing a robust real-time visual tracking algorithm is still very challenging.The challenge of visual object tracking is mainly due to the complexity and variability of tracking scenes(e.g.background clutter,occlusion)and the defects of the visual tracking algorithm.The difficulty of tracking algorithm design lies in the balance between accuracy and efficiency.This paper focuses on the real-time visual object tracking based on siamese network.The main contributions of this paper are as follows:(1)A tracking algorithm based on adaptive feature selection network is proposed(ASNT).In order to improve the ability of the general object tracking algorithm based on regression network to adapt to the target online,ASNT introduces a learnable selection unit to enhance the target-related convolution feature channel and reduce the noise response.By dividing the regression network into three components: feature representation,selection unit and motion regression,finely finetuning the network enhances the ability of the tracker to adapt to the target online.In addition,the channel dropout is proposed,which effectively avoids the over-fitting problem during offline training process and online learning.(2)A tracking algorithm based on classification and regression neural network and online patch filter network is proposed.In view of the fact that the siamese network is susceptible to similar interferences,an image mix-up method is used during offline training process,which alleviates the problem of imbalanced simple background and complex background dataset,and enhances the generalization ability of network trained with small scale dataset.During online tracking,an efficient patch filter network is proposed,which introduces the local detailed feature of the tracking object,and avoids the interference of similar objects to the tracker by means of online hard-negative learning,thus improving the overall tracking performance.The object tracking algorithm(ASNT and SiamRPN_OPF)proposed in this paper are verified by a large number of quantitative experimental analysis,attribute analysis,qualitative experimental analysis,module validity analysis,and parameter selection analysis on the authoritative dataset of object tracking and the performance of proposed algorithm is close to the state-of-the-art trackers.The algorithm proposed in this paper improves the accuracy of the tracking algorithm based on siamese network under the premise of ensuring real-time performance.The validity of the proposed tracking algorithm is proved by the module validity analysis experiment.