We found a match
Your institution may have rights to this item. Sign in to continue.
- Title
Taking full advantage of convolutional network for robust visual tracking.
- Authors
Guan, Hao; Cheng, Baozhong
- Abstract
Model-free visual object tracking has always been a challenging task in computer vision and multimedia analysis. How to alleviate the stability-plasticity dilemma plays a key role in making a robust tracker. In this paper, we propose a novel tracking framework which leverages convolutional network to tackle this fundamental problem. First, we construct a two-stream convolutional network which is pre-trained on a large number of sequential images. During tracking, hierarchical features of target are extracted by the network and an adaptive correlation classifier is trained on them to estimate the position change. In addition, the fully-connected layer with frozen weights of the network is deliberately designed as a self-correction agent which can re-detect the target in case of tracking failure. Through taking full advantage of the network by using both of its feature learning and fully-connected parts, the proposed tracking system can make a good balance between stability and plasticity. Experiments on large scale benchmark video sequences have shown that the proposed tracker outperforms many state-of-the-art tracking methods both in precision and robustness.
- Subjects
OBJECT tracking (Computer vision); ARTIFICIAL neural networks; TRACKING &; trailing; VIDEOS; PRECISION (Information retrieval)
- Publication
Multimedia Tools & Applications, 2019, Vol 78, Issue 8, p11011
- ISSN
1380-7501
- Publication type
Article
- DOI
10.1007/s11042-018-6679-9