In order to solve the problems of large scale variation and low resolution of tracking targets caused by fast movement,a Siamese network target tracking algorithm based on region refinement is proposed.A multi-scale feature sensing model is introduced into Siamese network to effectively extract the deep layer global channel features and local spatial features,so that the algorithm can extract discriminative information accurately.In order to further highlight foreground in the search area,a regional refinement model is constructed,target area features extracted by the backbone network are used to distinguish the search area targets,a coarse to fine tracking strategy is realized,effectively enhance the target representation ability.The proposed algorithm is tested against some existing tracking algorithms on OTB100 datasets.The experimental results show that the proposed algorithm has good performance in tracking success rate and tracking precision.At the same time,it shows strong robustness in low resolution,deformation and illumination variation.
关键词
深度学习/区域细化/目标跟踪/Siamese网络
Key words
deep learning/region refinement/target tracking/Siamese network
维普
万方数据