Object tracking algorithm based on fusion of SiamFC and Feature Pyramid Network
Zhongliang Wei, Yang Xiaohui · 2021 International Conference on Internet, Education and Information Technology (IEIT) · 2021
Aiming at the multi-scale problem of full convolutional siamese network SiamFC, this paper proposes a target tracking algorithm of siamese network based on the fusion of feature pyramid network. In the network template and detection branch, the up-sampling feature fusion method similar to feature pyramid network is adopted to enable the algorithm to track on multiple scales and improve the network's discrimination of small target objects. At the same time, VggNet-16 network is used to extract the shallow and deep features of the image, and adaptive fusion of the two features. Experimental results on OTB2015 and VOT2018 datasets show that, compared with SiamFC, the algorithm proposed in this paper can deal with scale changes and target drift, achieve better accuracy and success rate, and improve the speed of model tracking.