Siamese Network Object Tracking Algorithm Based on Squeeze-and-Excitation

Object tracking method based on Siamese Network has shown great performance, which expressed the tracking problem as dependency relationship between target template x and search template z. How to extract the characteristics of the target is a crucial problem. We explicitly model channels interdependencies within modules and adaptively calibrate the feature response between channels to enhance the representation of the entire network. To enable the network to perform feature recalibration, global information is learned to selectively enhance useful features and suppress few less useful features. We proposed a Siamese Network tracker based on SE-ResNet-50. In the training phase, meta-learning was introduced to solve the problem of lack of training data, so that the model can achieve a better result on an extremely small amount of data. Our method overcame the limitations of Siamese Network through weighted cross-correlation. Experimental results illustrate our algorithm is much better than some popular trackers in solving the problems of occlusion, illumination change and object deformation.