Siamese Network with Feature Fusion for Visual Tracking
Da Li, Yabing Kang, Xing Xiang, Wensheng Tao, Jiwei Hu · 2022 IEEE 6th Information Technology and Mechatronics Engineering Conference (ITOEC) · 2022
The visual tracking based on deep learning has developed greatly in recent years and has great advantages in accuracy and real-time speed. The target features are extracted through the backbone network. However, most networks only use the features extracted from the last layer of the network as output. In some complicated networks, the final features are mostly semantic feature information. The shallow appearance information is ignored. This greatly reduces the effective use of extracted feature information. In this study, a non-linear multi-layer feature fusion network combined with Siamese network is proposed to improve it. It uses CNN to fuse multi-layer features in a non-linear manner. The utilization of feature information and tracker performance indicators have been significantly improved. The new tracker is evaluated on the VOT2016 and UAV123 benchmarks. The result shows that the new tracker achieves very good performance and its performance has been significantly improved compared to the previous ones without the fusion network. The EAO is increased by 1.9% on VOT2016.