结合移动对象发现和对比学习的无监督跟踪OA北大核心

The conventional deep learning tracking methods require numerous manually annotated video labels to com-plete supervised learning for object tracking tasks,while unsupervised tracking methods can train models by using unla-beled videos,which is beneficial for deployment in real-world scenarios.Traditional unsupervised tracking methods heavily rely on exploring the spatial information of training samples,so it is difficult to track the target objects which have intense movement changes over a long period of time.This paper proposes an unsupervised tracking method based on the theory of cyclic consistency,which adopts unsupervised optical flow estimation and dynamic programming to discover the moving targets.In addition,for the purpose of utilizing the rich temporal information,a cyclic memory learning scheme is used to construct a memory queue,and a method based on contrastive learning is proposed to update the template.The experi-mental results show that the proposed tracking method achieves EAO of 0.402 and 0.344 on the VOT2016 and VOT2018 datasets respectively.The main performance indicators are comparable to some mainstream supervised learning methods in terms of tracking performance.