Journal
COMPUTER VISION AND IMAGE UNDERSTANDING
Volume 212, Issue -, Pages -Publisher
ACADEMIC PRESS INC ELSEVIER SCIENCE
DOI: 10.1016/j.cviu.2021.103272
Keywords
Correlation filter; Subspace learning; Subspace reconstruction; Visual tracking
Funding
- Funds for the National Natural Science Foundation of China [41371339]
- National R&D Program for Major Research Instruments of Natural Science Foundation of China [62027808]
- Fundamental Research Funds for the Central Universities [2017KFYXJJ179]
Ask authors/readers for more resources
The correlation filter achieves high performance in dealing with motion blur or lighting changes, but faces challenges in scenarios with occlusion. This study proposes the subspace reconstruction based CF (SRBCF) tracker to address the issue of background interference, achieving significant improvements in tracking accuracy.
The correlation filter (CF) achieves excellent performance, showing high robustness to motion blur or illumination change by learning filters. However, tracking in challenging scenarios with occlusion or out of-view is still not well resolved. In the scenario of occlusion, the background information is mixed into the image patch to learn the filter, which causes the filter to learn the background. To alleviate this problem, we improve CF trackers by proposing the subspace reconstruction based CF (SRBCF) tracker. In our method, the original image patch for learning filters is replaced by a reconstructed patch when the appearance of the object dramatically changes, such as occlusion or disappearance, so that the filter can learn from the object instead of the background. We construct the subspace with image patches of the searching window in previous frames. To track the changes in the subspace and mitigate the adverse effects of outliers on the subspace during the tracking process, we improve a dynamic L1-PCA method to construct and update the subspace with a slightly extra computational cost. Our method can be embedded in various correlation filter trackers, such as STAPLE and KCF. Extensive experiments on the OTB-100 dataset, UAV123, DTB70, and Temple Color Pure dataset (we removed 49 sequences repeated in OTB-100) validate the effectiveness of our method. The maximum AUC increase reaches 11.2% for the baseline method on DTB70.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available