SiamON: Siamese Tracking by Online Update Module and Non-local Pyramid Module

Abstract

The core idea of the object tracker based on Siamese network and its derivative is to transform the tracking problem into a template matching problem, cross-correlation the extracted object features and the convolution features of the search area, and achieve a good balance in tracking speed and accuracy. However, with the deepening of feature extraction network, the feature map contains less details, which is not conducive to object location. Secondly, the Siamese network can not adapt to the change of the object and obtain the context characteristics of the object, which makes the tracking effect worse when the object is occluded and deformed. In order to solve the above problems, we propose a non-local Siamese network–SiamON. Firstly, the non-local pyramid module is used to fuse the low-level and high-level network features, so that the convoluted features contain more details and semantic information, and enhance the object representation. Secondly, by adding an online update module, the error accumulation caused by template update is avoided, and the ability of the tracker to adapt to object changes online is improved, so that the tracker has strong robustness in object emission deformation and occlusion. Finally, extensive experiments are carried out on seven tracking datasets. The experimental results show that our tracker is better than the latest tracker and can run in real time.