Research on target tracking problem of fixed scene video surveillance based on unlabeled data

Qiyue Sun, Yuchen Wang, Yang Yang, Peng Xu

Abstract: Aiming at the problem that there are enough data in target tracking based on video surveillance, but the number of annotations is seriously insufficient or the quality is not high, this paper proposes a solution to the problem of target tracking based on unlabeled data. The target detection network with pre-trained weights is used to perform target detection on the objects in the unlabeled video. When the target is detected and the confidence is higher than 95%, the frame is extracted as the "first frame" of the KCF tracker. The KCF tracker will provide a large amount of labeled data as a training set to realize the training of the target tracking network MDNet. This paper achieves effective target tracking in fixed scene video surveillance based on unlabeled data and verifies that adding manually labeled data to the video surveillance data set will further improve network performance and prove the superiority and achievability of the proposed method.