三次元影像之追蹤與辨識法則研發

Abstract

隨著雷射測距儀的不斷發展，雷射測距的技術被使用在許不同的領域，像是距離量測、地形地貌 監測、三維立體物的量測等等。藉由三維雷射影像掃瞄儀能取得物體的距離影像以及強度影像，可以 提供物體的三維度空間的資訊。相較於傳統利用立體視覺影像產生的深度資訊，三維雷射影像掃瞄儀 除了有不需要校正向機參數的優點外，還能確保量測的結果相當精確。由於這些特點，以三維雷射影 像掃瞄儀所獲得物體的表面，可以被用來發展一套三次元影像識別的技術。建立三次元影像識別的技 術，首先要解決物體在不同視角下，產生不同型態表面輪廓的問題。換句話說，如何抽取出適當的特 徵，使得不論物體的姿態為何，所得到的特徵都不會改變，才能使辨識的工作順利完成。 本計畫發展三次元影像辨識法則的目標為：(1)建立與姿態無關的物體特徵，(2)在3-D 場景中偵測 感興趣的區域(ROI)，(3)自動辨識三維物體的演算法。在建立與姿態無關的物體特徵方面，引入旋轉 影像的概念，可以解決物體的特徵影像不隨著姿態改變而改變。由於旋轉影像是使用局部座標對物體 之表面影像點來編碼，所以旋轉影像的特徵並不會因為物體在三維空間的姿態而改變，基於這種特 性，本計畫中目標物體模型的特徵影像都是依照旋轉影像的概念來產生。在找出特徵影像後，利用3-D 線索法在一個大區域的場景來偵測出有興趣的區域(ROI)，再使用相關程序找出ROI 中與模型相匹配 的點並計算其可能的轉換，最後可運用遞迴最近點法則(iterative closest point algorithm, ICP)來驗證，得 到的結果為場景對應模式比較之比對辨識優適度，愈高的比對辨識優適度表示愈高的模式與實景之比 對信心指數，這種方法即是本計畫所使用的三維自動目標偵測與識別法則。

As the laser measurement system is developed continuously, laser range finder technology is used in many domains, like distance measurement, terrain monitor and 3-D object measurement. Both the range image and the intensity image of object will be acquired by the 3-D image laser scanner. The range image can provide 3-D information about object. Compares with depth information which is produced by the traditional stereoscopic vision image, the advantages of using 3-D image laser scanner are avoiding camera calibration and obtaining accurate result. Hence, the 3-D surface contour is obtained by 3-D image laser scanner can be used for 3-D object recognition. In order to develop an algorithm of 3-D object recognition, the most important thing in this project is to extract a pose-independent feature. The purposes of this 3-D object recognition project are: (1)pose-independent feature extraction, (2)ROI detection in a 3-D scene, and (3)automatic 3-D object recognition algorithm. First of all, the concept of spin image is introduced in pose-independent feature extraction. The spin image uses local coordinate system to encode all points on the surface of object. In addition, the feature is extracted by spin image cannot vary with different 3-D poses. In short, all pose-independent features of 3-D models in this project are generated based on this characteristic. After that, a 3-D cueing is employed to detect ROIs in a 3-D scene. This can help to find interested objects in a large-scale scene. Then use correlation process to match points of the model among ROIs and compute plausible transformation. Finally, the iterative closest point(ICP) algorithm is applied to verify transformation and to obtain the value of recognition goodness of fit. As a result, the higher value of recognition goodness of fit, the higher confidence index of scene-to-model is. Indeed, this is the automatic 3-D object recognition approach in this project.