| 题名: | An on-line learning tracking of non-rigid target combining multipleinstance boosting and level set |
| 作者: | Chen, Mingming1,2; Cai, Jingju1
|
| 出版日期: | 2013
|
| 会议名称: | Proceedings of SPIE: MIPPR 2013: Automatic Target Recognition and Navigation
|
| 会议日期: | 2013
|
| 学科分类: | Automatic target recognition - Image segmentation - Learning algorithms
|
| DOI: | 10.1117/12.2031170
|
| 中文摘要: | Visual tracking algorithms based on online boosting generally use a rectangular bounding box to represent the position of the target, while actually the shape of the target is always irregular. This will cause the classifier to learn the features of the non-target parts in the rectangle region, thereby the performance of the classifier is reduced, and drift would happen. To avoid the limitations of the bounding-box, we propose a novel tracking-by-detection algorithm involving the level set segmentation, which ensures the classifier only learn the features of the real target area in the tracking box. Because the shape of the target only changes a little between two adjacent frames and the current level set algorithm can avoid the re-initialization of the signed distance function, it only takes a few iterations to converge to the position of the target contour in the next frame. We also make some improvement on the level set energy function so that the zero level set would have less possible to converge to the false contour. In addition, we use gradient boost to improve the original multi-instance learning (MIL) algorithm like the WMILtracker, which greatly speed up the tracker. Our algorithm outperforms the original MILtracker both on speed and precision. Compared with the WMILtracker, our algorithm runs at a almost same speed, but we can avoid the drift caused by background learning, so the precision is better. © 2013 SPIE. |
| 英文摘要: | Visual tracking algorithms based on online boosting generally use a rectangular bounding box to represent the position of the target, while actually the shape of the target is always irregular. This will cause the classifier to learn the features of the non-target parts in the rectangle region, thereby the performance of the classifier is reduced, and drift would happen. To avoid the limitations of the bounding-box, we propose a novel tracking-by-detection algorithm involving the level set segmentation, which ensures the classifier only learn the features of the real target area in the tracking box. Because the shape of the target only changes a little between two adjacent frames and the current level set algorithm can avoid the re-initialization of the signed distance function, it only takes a few iterations to converge to the position of the target contour in the next frame. We also make some improvement on the level set energy function so that the zero level set would have less possible to converge to the false contour. In addition, we use gradient boost to improve the original multi-instance learning (MIL) algorithm like the WMILtracker, which greatly speed up the tracker. Our algorithm outperforms the original MILtracker both on speed and precision. Compared with the WMILtracker, our algorithm runs at a almost same speed, but we can avoid the drift caused by background learning, so the precision is better. © 2013 SPIE. |
| 收录类别: | Ei
|
| 语种: | 英语
|
| 卷号: | 8918
|
| ISSN号: | 0277786X
|
| 文章类型: | 会议论文
|
| 页码: | 891802
|
| Citation statistics: |
|
| 内容类型: | 会议论文
|
| URI标识: | http://ir.ioe.ac.cn/handle/181551/7700
|
| Appears in Collections: | 光电探测与信号处理研究室（五室）_会议论文
|
| File Name/ File Size |
Content Type |
Version |
Access |
License |
|
|---|
| 2013-2099.pdf（589KB） | 会议论文 | -- | 限制开放 | | View
联系获取全文 |
|
|
作者单位: | 1. Institute of Optics and Electronics, Chinese Academy of Sciences, Chengdu, Sichuan Province, 610209, China
2. Graduate University of Chinese Academy of Sciences, Beijing, 100049, China
|
| Recommended Citation: | |
Chen, Mingming,Cai, Jingju. An on-line learning tracking of non-rigid target combining multipleinstance boosting and level set[C]. 见:Proceedings of SPIE: MIPPR 2013: Automatic Target Recognition and Navigation. 2013.
|
|
|
