Research On Visual Object Tracking Based On Multi-layer Feature Fusion And Nonlinear Template Updating

Video object tracking is a research hotspot in the field of computer vision in recent years,which has attracted the interest of the majority of researchers.It has a wide range of applications in visual surveillance,human-computer interaction,and virtual reality.However,there are still some deficiencies and challenges in the current object tracking algorithm.The algorithm model with high accuracy and strong robustness has become the goal pursued by researchers.Object tracking algorithms based on deep learning have received extensive attention in recent years,but most of the model algorithms cannot make full use of the extracted feature information and ignore the model update strategy often used in traditional filtering.It causes that the trackers can not obtain better performance in accuracy and robustness.Based on the Siamese Network tracking algorithm,this paper conducts in-depth research on the nonlinear template update strategy based on multi-layer feature fusion.And different feature extraction networks are used to study the impact on the above two points.The main research contents are as follows:(1)Aiming at the problem of insufficient utilization of feature information in the current tracking model based on deep learning,a multi-layer feature fusion method is introduced into the feature extraction network to increase the utilization of deep and shallow features.Through extracting features from different layers at the same time and cropping these extracted features to the same size,the convolutional neural network is used to fuse them.The fused features contain shallow appearance information and deep semantic information,which becomes more discriminative.It increases the reliability of similarity comparison.Due to the increase in accuracy,it provides a basis for subsequent model updates.(2)Aiming at the unreliable template caused by object deformation and occlusion in the current deep learning-based tracking model,a nonlinear update strategy based on convolutional neural network is introduced to increase the accuracy and robustness of the algorithm.In general,the linear template update in the Siamese network will cause the template information to decay exponentially with time,but the nonlinear model update strategy introduced is to let the tracker learn the update of the template by itself.And each template is a cumulative template,including the template information from the start of tracking to the current one.(3)Research on the performance improvement of deep network for feature fusion and nonlinear template update.Aiming at the problem that the features extracted by different networks have different performances of the tracker,a deep network is introduced for feature extraction to study the impact of the deep network on the performance of the tracker proposed in this study.And on the basis of feature fusion and model update,the difference of feature extraction effect between deep network and shallow network is compared and studied.