Two-stage Object Tracking Based on Similarity Measurement for Fused Features of Positive and Negative Samples

Abstract

AbstractThe tracking algorithm based on Siamese networks cannot change the corresponding templates according to appearance changes of targets. Therefore, taking convolution as a similarity measure finds it difficult to collect background information and discriminate background interferents similar to templates, showing poor tracking robustness. In view of this problem, a two-stage tracking algorithm based on the similarity measure for fused features of positive and negative samples is proposed. In accordance with positive and negative sample libraries established online, a discriminator based on measurement for fused features of positive and negative samples is learned to quadratically discriminate a candidate box of hard sample frames. The tracking accuracy and success rate of the algorithm in the OTB2015 benchmark dataset separately reach 92.4% and 70.7%. In the VOT2018 dataset, the algorithm improves the accuracy by nearly 0.2%, robustness by 4.0% and expected average overlap (EAO) by 2.0% compared with the benchmark network SiamRPN++. In terms of the LaSOT dataset, the algorithm is superior to all algorithms compared. Compared with the basic network, its success rate increases by nearly 3.0%, and the accuracy rises by more than 1.0%. Conclusions: The experimental results in the OTB2015, VOT2018 and LaSOT datasets show that the proposed method has a great improvement in the tracking success rate and robustness compared with algorithms based on Siamese networks and particularly, it performs excellently in the LaSOT dataset with a long sequence, occlusion and large appearance changes.KeywordsObject trackingSiamese networkQuality evaluationOnline update of sample librarySimilarity measure