即時目標物追蹤系統設計與實作

Design and implementation of a real-time object tracking system

馮振東(Chen-Tung Feng)

指導教授 ： 陳永盛

元智大學/工程學院/電機工程學系/碩士(2004年)

摘要

不論是在哪一個場合，例如電視轉播，保全監控系統，甚至是家庭娛樂之上，即時目標追蹤都是一個非常重要的課題。一個良好的即時追蹤系統最重要的是能夠快速而準確的定位到目標，而且能夠主動的鎖定目標，以供我們後續研究分析。但是能夠同時兼上述的條件追蹤系統是非常不容易建構的，之前眾人所提出的追蹤系統都難以同時達到以上的條件，不是無法兼顧其中一項特性降低複雜度以換取系統實作，就是需要昂貴的硬體來提升效能。因此，近年來不斷有人提出新的架構來克服此問題。
本文提出一個以核心導向物件追蹤為基本核心的即時目標物追蹤改良系統，其目標模型是用機率密度函式所構成，再以平均位移法來求取新的候選區。本篇論文主要分為三個部份：第一、我們介紹核心導向物件追蹤的基本原理及概念。我們採用其演算法，因為其方法能同時兼顧了速度及準確率。第二、核心導向物件追蹤的主架構預留了許多的空間，能讓我們能進一步的改良與修正，用以提升此演算法的效能及正確性，例如加入卡門濾波器，動態移動機制等等。最後是實驗的結果展示與討論。由我們的實驗分析證明，我們所提出的改良方法有極佳的穩定性與實用度。我們的系統能在每秒30個畫面的速度下運作，每個畫面能在平均5次遞迴運算之內，計算出目標的位移，正確的執行即時的追蹤。這證明了這追蹤系統足以達到我們的需求。

關鍵字

即時追蹤 ； 平均位移 ； 卡門濾波器 ； 巴達查里亞係數 ； 目標定位

並列摘要

Real-time tracking system is a very important study whenever what occasions we are, for example, TV broadcasting, security system, entertainments, and so on. It is the most important thing that a good real-time tracking system must be able to locate the target fast and correctly, and can lock on the target voluntarily for the following analyses. But it is difficult to build a tracking system which gives consideration to the conditions we list above. The methods that people bring up are hard to reach those conditions, either it has to throw away one of the conditions to reduce the complexity, or needs the expansive equipments to increase performance. Therefore, in recent years, many people bring up their idea to solve this question.
In this thesis, a real-time object tracking system is developed based on the kernel-based object tracking, which its model is based on PDF, and uses mean-shift to locate the new candidates. This thesis consists of three main parts. First, the basic idea and principle of a kernel-based object tracking is reviewed. We adopt this algorithm for improvement because this method can give consideration to both speed and accuracy. Secondly, the main structure of kernel-based object tracking has reserved some space for us to modify further, to increase the tracking efficiency and correctness. For example we can embed Kalman filter and dynamic movement to improve the system. Finally, we show some experiments and give a discussion. Our experiments have confirmed that the proposed tracking system performs well stability. Our system can track the target at the speed of 30 frames per second; the average number of iterations for target tracking in every frame is less than 5. It proves that it is a nice system which can satisfy our requirements.

並列關鍵字

real-time tracking ； mean-shift ； Kalman filter ； Bhattacharyya coefficient ； target localization

參考文獻

[1] Cucchiara, R., Grana, C., Piccardi, M., Prati, A., “Detecting moving objects, ghosts, and shadows in video streams”, IEEE Transactions on, Pattern Analysis and Machine Intelligence, Volume: 25, Issue: 10, Oct. 2003, Pages: 1337 — 1342.

[2] Szeliski, R., “Video mosaics for virtual environments”, IEEE, Computer Graphics and Applications, Volume: 16, Issue: 2, March 1996, Pages: 22 — 30.

[3] Durucan, E., Ebrahimi, T.,”Change detection and background extraction by linear algebra”, Proceedings of the IEEE, Volume: 89, Issue: 10, Oct. 2001, Pages: 1368 — 1381.

[4] Duncan, J.H.; Chou, T.-C,”On the detection of motion and the computation of optical flow”, IEEE Transactions on, Pattern Analysis and Machine Intelligence, Volume: 14, Issue: 3, March 1992, Pages: 346 — 352.

[5] Malo, J., Guttierrez, J., Epifanio, I., Ferri, F.J,” Perceptually weighted optical flow for motion-based segmentation in MPEG-4 paradigm”, Electronics Letters, Volume: 36, Issue: 20, 28 Sep 2000, Pages: 1693 — 1694.