| Object tracking is an important research focus of computer vision.Its main task is to track a moving target given in the initial frame of a video,and perform state estimation and label in the subsequent frames.Object tracking not only expands rich data resources for scene analysis tasks,but also provides plentiful help for correct detection and identification of targets.It plays a key role in quiet a few fields such as intelligent monitoring,medical diagnosis,national defense and security,smart city and automatic driving.With the rapid development of deep learning-related technologies in the field of object tracking,object tracking algorithms that use convolutional neural networks and Siamese neural networks to promote tracking performance have achieved remarkable results and become mainstream research methods.Nevertheless,in practical application scenarios,the tracker will be affected by the external environment and the internal factors of the object,consequently it is still challenging to attain accurate and robust object tracking tasks.Aiming at the challenging issue that affects the performance of the tracker,this paper proposes starts from optimizing the feature extraction network structure and improving the algorithm model,and proposes two target tracking algorithms based on Siamese neural networks.For the first tracking method,a Siamese network tracking algorithm based on channel attention and Kalman filter is proposed for the tracking drift problem lead to by occlusion and fast-moving objects.Firstly,the channel attention structure is embedded in the Siamese network model,and the correlation among the channels in the image is analyzed to improve the representation ability of the network to extract features.Subsequently the Kalman Filter is introduced into the Siamese network to obtain the Spatiotemporal trajectory information of the object,and the state of the object tracking is judged by a high confidence method based on the Average Peak-toCorrelation Energy,which makes the object tracking algorithm more robust and improves the success rate of the tracking.For the second tracking method,a Siamese network tracking algorithm based on frequency channel attention and template adaptation is proposed for the situation that the object scales change and the influence reduce the tracking accuracy caused by the size change of target and the influence of similar objects lead to the decrease of tracking accuracy.The tracking algorithm uses frequency channel attention network for sampling,and introduces more frequency components to make full use of feature information,which makes up for the shortage of existing channel attention methods to extract feature information,enhances the identification of similar features,and improves the appearance of similar targets.At the same time,in order to better cope with the tracking objects of different scales,the scales of the object template are adaptively selected according to the ratio of the object and the background,which improves the accuracy of the tracking algorithm.Experiment show that the method balances the robustness and speed of tracking well.In addition,a large number of comparative experiments with different tracking algorithms are carried out on seven benchmark datasets,and a variety of mainstream evaluation methods are used for qualitative and quantitative evaluation and theoretical analysis,which proves that the algorithm proposed in this paper has strong robustness. |
