基于稀疏稠密結(jié)構(gòu)表示與在線魯棒字典學(xué)習(xí)的視覺跟蹤
doi: 10.11999/JEIT140507 cstr: 32379.14.JEIT140507
基金項目:
國家自然科學(xué)基金(61175035, 61379105)資助課題
Visual Tracking Based on Sparse Dense Structure Representation and Online Robust Dictionary Learning
-
摘要: L1跟蹤對適度的遮擋具有魯棒性，但是存在速度慢和易產(chǎn)生模型漂移的不足。為了解決上述兩個問題，該文首先提出一種基于稀疏稠密結(jié)構(gòu)的魯棒表示模型。該模型對目標模板系數(shù)和小模板系數(shù)分別進行L2范數(shù)和L1范數(shù)正則化增強了對離群模板的魯棒性。為了提高目標跟蹤速度，基于塊坐標優(yōu)化原理，用嶺回歸和軟閾值操作建立了該模型的快速算法。其次，為降低模型漂移的發(fā)生，該文提出一種在線魯棒的字典學(xué)習(xí)算法用于模板更新。在粒子濾波框架下，用該表示模型和字典學(xué)習(xí)算法實現(xiàn)了魯棒快速的跟蹤方法。在多個具有挑戰(zhàn)性的圖像序列上的實驗結(jié)果表明：與現(xiàn)有跟蹤方法相比，所提跟蹤方法具有較優(yōu)的跟蹤性能。
-
關(guān)鍵詞:
- 視覺跟蹤 /
- 稀疏表示 /
- 稠密表示 /
- 字典學(xué)習(xí)
Abstract: The L1 trackers are robust to moderate occlusion. However, the L1 trackers are very computationally expensive and prone to model drift. To deal with these problems, firstly, a robust representation model is proposed based on sparse dense structure. The tracking robustness is improved by adding an L2 norm regularization on the coefficients associated with the target templates and L1 norm regularization on the coefficients associated with the trivial templates. To accelerate object tracking, a block coordinate optimization theory based fast numerical algorithm for the proposed representation model is designed via the ridge regression and the soft shrinkage operator. Secondly, to avoid model drift, an online robust dictionary learning algorithm is proposed for template update. Robust fast visual tracker is achieved via the proposed representation model and dictionary learning algorithm in particle filter framework. The experimental results on several challenging image sequences show that the proposed method has better performance than the state-of-the-art tracker.-
Key words:
- Visual tracking /
- Sparse representation /
- Dense representation /
- Dictionary learning
-
計量
- 文章訪問數(shù): 2501
- HTML全文瀏覽量: 167
- PDF下載量: 843
- 被引次數(shù): 0
下載:
