Template pool updating and feature combination object tracking

Abstract

In the target tracking task, most Siamese network target tracking algorithms only use the template information of the first frame and do not make good use of the feature information of the subsequent frames, which shows poor robustness and target drift in dealing with target occlusion and appearance change in complex scenes. In this regard, an online template pool update module is established based on the SiamFC++ algorithm to store the information of reliable template frames, and an accuracy and necessity judgment mechanism of template update is proposed to improve the reliability of the final template. In addition, a feature combination module is proposed to capture the information of different scale regions and obtain more scale information through an asymmetric dilated convolution layer. Experiments and analysis are conducted on OTB2015 and VOT2018 datasets, and the proposed algorithm improves the accuracy and success rate by 1.7% and 1.7%, respectively, on OTB2015 compared with the SiamFC++ algorithm; the accuracy and average overlap rate improve by 0.8% and 6.1%, respectively, on VOT2018, and the robustness decreases from 0.169 to 0.140, and the tracker more stable.