Object Tracker Based on Siamese Network with Displacement Penalty and Template Update

Chenyu Huang,Yu Zhang,Yi Zhao,Heng Dong,Zhuoming Li

DOI: https://doi.org/10.1109/HDIS56859.2022.9991726

2022-01-01

Abstract:SiamFC is equipped with a fully-convolutional Siamese network trained end-to-end on the ILSVRC15. Despite the success of the algorithm, it's still susceptible to interference from background clutter. In order to solve the problem, we propose an object tracking algorithm based on SiamFC. We introduce a spatial displacement penalty module into the tracker, to spatially weight the position where multiple peaks may occur on the score map. Using the correlation between two consecutive frames, tracker can reduce the probability of tracking the wrong analogues. SiamFC is difficult to track the object after deformation, so we define the confidence degree to measure the trustworthiness of tracking results. When the degree reaches the threshold, tracker updates the object appearance model selectively. We exploit the spatial and temporal information of the object and background to optimize the tracking process. Our tracker has higher robustness compared to SiamFC algorithm with the same complexity. We test our tracker on TC128 and VOT dataset, the simulation results show that our tracker achieves better performance in precision, AUPC and IoU than SiamFC.