基于难样本挖掘的孪生网络目标跟踪

In complex environment, the object tracking algorithm of fully-convolutional siamese network is prone to track drift or even track failure. In order to solve the problem, this paper proposed a siamese network tracking algorithm based on hard sample mining. On the basis of SiamFC, this method first used a feature fusion module for feature fusion to enhance the robustness of feature representation, and then proposed a novel loss function to strengthen the learning ability of network to hard samples and alleviate the problem of imbalance between positive and negative samples. To verify the validity, this method was tested on OTB2015 benchmark and GOT10k dataset. The results of OTB2015 show that this method increases the success rate by 2.6% and the accuracy by 2% compared with SiamFC. On the GOT10k dataset, the mAO of this method is 0.429, which is 3.7% higher than the SiamFC. It illustrates that this method has a better performance in the case of illumination variation, object deformation, and similar background interference.