即時化自適性提昇演算法應用於物件追蹤之改良與延伸

The Evaluation and Extension of On-line Boosting for Object Tracking

張敦傑(Tun-Chieh Chang)

指導教授 ： 莊永裕

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2009年)

https://doi.org/10.6342/NTU.2009.03377

摘要

本論文提出一個改善即時化自適性提昇演算法(On-line Boosting) 應用於物件追蹤的方法，藉由與物體生成模型的有效連結，使原先可能產生的定位偏差減少，進而提昇追蹤系統的準確與穩定性。即時化自適性提昇演算法是一種分類式的學習模型，他透過統計與分析目標物和背景物的特徵，建立出判斷邊界並能迅速地做出分類。這種模型擅長即時的學習和適應，並且在傳統的分類問題上具有很好的區別能力。但在極富挑戰性的物件追蹤問題上，由於追蹤物外觀的多樣變化，一旦模型出現微小的誤差，錯誤便會持續被學習並累積，導致追蹤的失敗。因此我們希望在追蹤系統中，加入生成式模型，以提昇追蹤的準確性。生成式物體模型可視為一種描寫物體外觀的方法，它會針對目標物作複雜度較高的敘述，而這樣的模型常是物體辨識的基礎。我們的追蹤演算法架構，試圖結合較為精細但運算耗時的生成式模型，補強即時化自適性提昇演算法的弱點，並運用重點取樣的原則，保留原本分類式模型具有的彈性和速度。我們透過對公開測試影片和實際應用中影片的衡量，比較原始演算法與改良後系統的平均誤差量，結果顯示我們確能有效地結合兩種不同原理的物件學習模型演算法，在當中取得平衡且使其互相支援，進而使追蹤系統得到更好的效能。

關鍵字

物件追蹤 ； 即時化自適性提昇

並列摘要

In this thesis, we present an extension of On-line Boosting on tracking problem which enhances the performance of the tracking system. We want to reduce the errors during the tracking process, and improve the accuracy and the robustness of the tracker. On-line Boosting is a discriminative-based training model, and it has the ability to fast distinguish the positive from the negative by analysing the distributions of both the target and background to build the decision boundary. This kind of learning model provides good adaptivity, and its ability is remarkable in the typical classification problems. But when the task becomes the challenging tracking problem, the variety of appearance changes of an object makes the model easily generating slight errors. The error would propagate through the tracking process and the accumulated errors cause tracking failure. For this reason, we want to combine the On-line Boosting with a more precise generative-based training model. Generative-based model can be viewed as a description of the object appearance, and it represents the subject in a more complex way. This kind of models usually play an important role of object recognition. Our tracking framework aims to take advantages of the slow but precise generative models, to compensate for the defects of On-line Boosting in tracking algorithm. Even more, we link two models by importance sampling and that retains the speed and the adaptivity of discriminative models. We evaluate our system by estimating the average position error in sequences, and we try our method on public testing data and the real-world data. The result shows that we can effectively combine two object training models with different concepts, making a balance collaboration between these methods, and finally improving the performance of the tracking system.

並列關鍵字

Object tracking ； On-line Boosting

參考文獻

[1] S. Avidan. Support vector tracking. In IEEE Trans. on Pattern Analysis and Machine

Intelligence, pages 184–191, 2001.

[2] S. Avidan. Ensemble tracking. In In CVPR, pages 494–501, 2005.

[3] B. Babenko, M.-H. Yang, and S. Belongie. Visual Tracking with Online Multiple

Instance Learning. In CVPR, 2009.