植基於角點幾何關係之快速物件辨識與追蹤

Fast Object Recognition and Tracking based on Corner Geometric Relationship

古侑弘(Yu-Hung Ku)

指導教授 ： 陳金聖

國立臺北科技大學/機電學院/自動化科技研究所/碩士(2010年)

https://doi.org/10.6841/NTUT.2010.00376

摘要

本論文提出一種可快速執行物件辨識之新計算結構，及一種可準確追蹤運動物體的伺服追蹤系統。
一般物件辨識都是利用SIFT-PCA的角點特徵描述來做匹配，再利用RANSAC法則去除錯誤的匹配點，這樣的做法將會非常的耗時。在本論文中，我們將重新組合SIFT-PCA與RANSAC的順序流程，並且緊密結合角點幾何關係來挑選強健性足夠的初始配對點。利用這些初始配對點，可找出待測影像與樣本影像之間的轉換矩陣，再由此轉換矩陣可找出待測影像與樣本影像之所有配對點組，而不需再作SIFT-PCA計算。所以本法則比起一般物件辨識法則快速許多。從實驗結果可看出，本法則只需花費一般物件辨識法則的六分之一到七分之一的時間。雖然，本法則之準度有些微的誤差，但仍然能保持在一個不錯的誤差水準範圍內。
當我們有興趣的物體進入我們的視野，且已被辨識出來，我們將希望跟隨追蹤它運動。在本論文，我們將一個攝影機座架設在以伺服馬達帶動的平台的伺服追蹤系統來達成。這個伺服追蹤系統追蹤物件時，能夠左右轉動，使物件盡可能一直維持出現在攝影機螢幕的中央。由於我們的系統能夠工作得遠較以傳統辨識法則辨識目標的系統快速，容易擁有較高的資料更新率與較接近理想的追蹤效果，從實驗結果可以驗證此項優點。本論文中也使用一alpha-beta濾波器，能夠過濾一些因測量雜訊或平台移動所產生的誤差，使得到較準確的資料，供未來各種運用。

關鍵字

物件識別 ； 角點偵測 ； 幾何關係 ； 物件追蹤

並列摘要

This thesis proposes a new computation architecture which can recognize the interesting object very fast. And, a servo tracking system is also proposed to track the moving object accurately.
The general object recognition algorithm which uses corner descriptions based on Scale Invariant Feature Transform (SIFT) and Principal Component Analysis (PCA) for patch matching and then eliminate the false matching pairs by Random Sample Consensus (RANSAC) would have very heavy computation load. In this paper, we re-arrange the operating sequence of SIFT, PCA and RANSAC and utilizes the corner geometry relationships to pick up the good initial pairs with robust property. Then, the parameters of the transformation matrix between the recognized image and template image can be solved and all matching pairs can be obtained by the transformation matrix without SIFT-PCA calculation. Only very few SIFT and PCA calculation is required in our algorithm so that it is much faster than the tradition algorithms.From the experimental results, our algorithm only spent around one sixth to seventh computation time comparing to traditional algorithm. In the other hand, our algorithm would have a little poor recognized error, but it is still in the level of good performance in general cases.
While the interesting object has been found and recognized, we will like to track this object wherever it moves. In this thesis, we have a servo tracking system which the camera mountained on a platform driven by the servo. This servo system can track the moving object by rotating left and right to keep the object always appears on the center of the screen of the camera approximately. Since our system can work much faster than the conventional system which uses the conventional algorithm to recognize the object, higher data update rate can easily be achieved and better performance can be anticipated in our system. The experimental results can confirm this advantage. An alpha-beta filter is also used in our system to get more accurate data for further applications. It will smooth out the errors originated from the measurement noises and platform moving effectively.

並列關鍵字

Object Recognition ； Corner Detection ； Geometric Relationship ； Tracking

參考文獻

[17] C. S. Chen, M. F. Tsai, and C. C. Chiu, " Object Recognition Based on Modified Intuitive Corner Detection and Two-stage Corner Matching," 2009 SICE Internatioal Conference, Fukuoka, JAPAN, Aug. 2009.

[25] Y. S. Shiao, "Design and Implementation of Real-Time Tracking System Based on Vision Servo Control," Tamkang Journal of Science and Engineering, Vol. 4, No. 1, pp.45-58, 2001.

[4] D. G. Lowe, "Distinctive image features from scale invariant keypoints," International Journal of Computer Vision, Vol. 60, No. 2, pp.91-110, 2004.

[7] I. T. Joliffe, “Principal component analysis,” New York: Springer-Verlag, 1986.

[8] K. Mikolajczyk and C. Schmid, "A performance evaluation of local descriptors," IEEE Transactions on Patten analysis and Machine Intelligence, Vol. 27, No. 10, pp.1615-1630, October 2005.