A Study Of Single Target Tracking Based On Siamese Network And Fusion Of Multi-scale Deep Feature

With the rapid development of computer vision,target tracking has made great progress,which has been used in many visual scenes,such as intelligent video surveillance,human-computer interaction,robot visual navigation,medical diagnosis and so on.Many target tracking algorithms represented by correlation filtering and deep learning show excellent performance,among which the method based on deep learning is dominant.Convolution neural networks are composed of several convolution layers,and the output of different convolution layers maps the different features of the image,where the deep network may extract more semantic features,while the shallow network may be more sensitive to the shape changes of objects.Therefore,how to integrate the multi-scale features obtained by different layers of network and give full play to the role of various features is a hot topic in the field of vision.To this end,this thesis designed two new single target tracking network models based on the traditional Siamese network system,aiming to optimize the feature extraction:(1)Single target tracking method based on feature pyramid.Feature pyramid is widely used in visual field because it can fuse the semantic information of different levels of image data.For this reason,this thesis uses a strategy of fusing multi-scale features,and the feature maps of the feature extraction backbone network called Alex Net are pooled into different scales to form a kind of pyramid.Then,the multi-scale feature maps in the feature pyramid are un-sampled into the same size as input and spliced with the input feature maps in turn,so as to realize fusion of multi-scale features.Finally,the fusion features are input into the region proposals networks for target localization.(2)Single target tracking method based on fusion of multi-scale feature and texture feature.Texture features play a very important role in many image processing applications,such as face recognition,image classification and so on.In the application of video tracking,the texture features of the target have robustness to some extent.Therefore,this thesis uses a local binary convolution network to extract the texture features of the target,and then realize the fusion of multi-scale features through the proposed new feature pyramid network.The feature fusion network proposed in this thesis adopts a residual learning strategy to realize the effective fusion of different scale features.Therefore,this model can learn the relationship between adjacent scale features more efficiently,so as to achieve better feature expression.The combination of local binary module and pyramid residual learning module greatly enhances the effect of image feature extraction.In order to realize the target location,the method also input the fused feature into the region proposals networks.For the purpose of validating the two ideas presented in this thesis,the open object tracking datasets(VOT2015,VOT2016,VOT2018 and OTB100)are tested and the testing results are compared.The experimental results show that the proposed method achieves better tracking performance(accuracy,success rate,etc.)than some previous trackers.