Template Update Based on 3D-Convolutional Siamese Network for Object Tracking

Abstract

Object tracking is an important research area in the field of computer vision. In the past two years, the object tracking algorithms based on Siamese network have yielded brilliant results in CVPR. However, previous algorithms only extract the object feature of the first frame as a tracking template. In the process of tracking the object, the object template remains unchanged, leading to poor tracking accuracy. In view of this, the present paper proposes a new, end-to-end trained, fully convolutional 3D Siamese network-based tracking algorithm to extract multiple features. Logistic loss function and SGD are used to train the network. The trained network realizes the use of multi-frame features to update the object template in the process of tracking the video's object. The tracker in this paper can run at real-time frame-rates in OTB, VOT, and GOT-10k. The algorithm is applied to SiamFC, its accuracy is improved by 4% and 3% on the OTB and VOT-2016 data sets, respectively.