視覺物體追蹤應用複合差異式信心融合方法

Visual Object Tracking Based on a Hybrid Discriminative Confidence Fusion Approach

高竟中(Ching-Chung Kao)

指導教授 ： 羅仁權

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2011年)

https://doi.org/10.6342/NTU.2011.10878

摘要

視覺物體追蹤在很多機器人和電腦視覺的應用中，常常是一個相當關鍵的步驟。這些應用包括視覺伺服、保全監看、擴增實境、人機互動等。本論文提出一個複合差異式的視覺物體追蹤演算法，此複合式演算法結合了兩個差異式追蹤演算法。差異式物體追蹤的概念就是把物體追蹤當成是分類問題，亦即把物體追蹤當成是分離目標物和背景的過程。本論文提出的方法藉由一組規則把兩個差異式追蹤演算法結合在一起，並且透過信心融合的方式來融合兩個差異式追蹤演算法的結果。
通常差異式追蹤演算法都會利用分類器來分類目標物和背景，因此本論文會討論一些學習演算法，包括監督式學習和半監督式學習的演算法。本論文也會比較了不同的學習演算法在視覺追蹤上的效果，其中最適合的兩種學習演算法，被拿來應用在所提出的追蹤演算法上。
這兩個不同的差異式追蹤演算法是被設計來互補彼此分辨目標物和背景的能力，所以最終的複合式演算法預期會比原本的兩個演算法都還要好。為了使兩個差異式追蹤演算法能互補分辨目標物和背景的能力，一個追蹤演算是在像素階層分離目標物和背景，而另一個是在區塊階層分離目標物和背景。此外，兩者採用不同的學習演算法訓練分類器，一個是兩元的分類器，另一個則是多類別的分類器。為了適應物體在追蹤過程中的外表變化，分類器會不斷的更新。最終的複合式視覺物體追蹤演算法會利用一組包括信心融合，還有追蹤器切換所組成的規則，來結合兩個差異式視覺物體追蹤器。
本論文所提出的方法會與其他三種現今流行的視覺物體追蹤演算法，在兩個十分有挑戰性的影像序列做比較，並且根據兩種不同的方式來評價結果的好壞。第一種評價方式是平均的追蹤誤差，而第二種方式則是根據重疊面積的多寡。實驗的結果顯示，本論文所提出的方法，在所有的測試中都是最好或者是第二好的。

關鍵字

視覺物體追蹤 ； 複合差異式方法 ； 電腦視覺 ； 智慧型機器人

並列摘要

Visual object tracking plays a key role in many robotics and computer vision applications, such as visual servoing, surveillance, augmented reality, and human-robot interaction. In this thesis, a hybrid visual object tracking algorithm that combines two discriminative trackers is proposed. Discriminative trackers model object tracking as a classification problem, that is, they try to distinguish targets from backgrounds. The proposed hybrid approach combine the two component trackers by a set of rules, and the tracking results of the two trackers are fused to determine the object location via a process called confidence fusion.
Incorporating one or more classifiers into the trackers is a convenient way to implement the concept of the discriminative object tracking. Therefore, several learning algorithms, including supervised and semi-supervised ones, are reviewed in this thesis, and the tracking results using different learning algorithms are compared in this thesis. The two most suitable learning algorithms, AdaBoost and Random Forests, are employed in the proposed approach to train the classifiers.
The hybrid tracker is hopefully more powerful since the two trackers could complement the ability of discrimination. To achieve this goal, one tracker extracts image features pixel by pixel, and the other extracts image features over several rectangular regions. In addition, the corresponding classifiers are trained using different learning algorithms, a binary-class learner and a multi-class learner, respectively. The classifiers can be updated online during tracking so that the trackers are adaptive to the variations of the appearance of the target. A set of rules including tracker switching and confidence fusion is devised to synthesize the two trackers.
The proposed algorithm is compared with three other popular tracking algorithms on two challenging video sequences, which include drastic appearance variations and occlusions. The tracking results are evaluated by two different ways. One is based on the center location error, and the other is based on the overlapping area. The experimental results show that the proposed approach is either the best or the second best among all the tests.

並列關鍵字

visual object tracking ； hybrid discriminative approach ； computer vision ； intelligent robot

參考文獻

[1] A. Adam, E. Rivlin, and I. Shimshoni, “Robust fragments-based tracking using the integral histogram,” IEEE Conference on Computer Vision and Pattern Recognition, 2006.

[2] S. Avidan, “Support vector tracking,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 26, pp. 1064–1072, 2004.

[3] S. Avidan, “Ensemble tracking,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 29, pp. 261–271, 2007.

[4] B. Babenko, M.-H. Yang, and S. Belongie, “Visual tracking with online multiple instance learning,” IEEE Conference on Computer Vision and Pattern Recognition, 2009.

[5] M. Black and A. Jepson, “EigenTracking: Robust matching and tracking of articulated objects using a view-based representation,” International Journal of Computer Vision, vol. 26, pp. 329–342, 1998.