Abstract
Three- dimensional object recognition is one of the most interesting andimportant research topices in computer vision. Though 2- D image analysisin some situations serves to capture the intrinsic nature of the problem,3- D models are essential in the analysis and interpretation ofarbitrarily shaped 3- D objects. This often entails the provision ofobject descriptions in 3- D space. Generally, two methods are available,one uses a ranging device and the other combines multiple views fromdifferent viewpoints. Either approach bears some drawbacks in efficiency,accuracy, or performance. In this dissertation, we propose the strategy ofintegrating intensity and range information to build representations ofobjects, to recognize them, and to locate the positional parameters forrobotic manipulation.We provide a categorization and discussions of fusion. "Sensorintegration" concerns the manner in which multiple, redundant, anddisparate sensors are cooperated to accomplish one work with the bestefficiency and accuracy and to reuce sensor errors. We propose animplementation of "intensity-guided range sensing" strategy in whichrange data are selectively sensed with the guidance of intensity features.Also, based upon their complementary natures, range data analysis aids inthe detection and recovery of intensity errors, such as missing orauperfluous edges after line-drawing extraction. On the other hand, "datafusion" disregards the physical sensor operation. It lays emphases on thecombination of information from multiple sensors, aiming at the computationof optimal results from a number of observations(passive fusion) orinference of outcomes that any single source of sensor data are hard orimpossible to obtain (active perception). We apply a range finder and acamera to capture the images from different viewpoints and then constructa fused description about the object. Range and intensity domains arefirstly registered via spatial transformation, by which range jump cluesare analyzed in the intensity domains to infer region adjacencies thatare initially missing in both individual domains, Examples show that thefused results convey much more abundant information for recognition.Either paradigm introduced above constructs an attributed relational graphrepresenting the unknown object to be recognized. Each node represents oneobject face and each arc specifies the geometrical and topologicalproperties between two faces. Recognition of 3- D objects is then reducedto a monomorphism matching problem. The "interpretation tree search"technique, as well as the operation of node ordering and null assignmentsas proposed, are adopted to perform node pairing, with the arc attributesacting as auxiliary constraints for optimal search. Since 3- D data areavailable, they can be utilized to compute a transformation (translationand rotation) between the model and sensor coordinate systems for furtherrobotic manipulation. A non- iterative method based upon the "quaternion"technique is adopted to accomplish the work. Experiments show that resultsof recognition and positioning are satisfactory.三度空間物體辨認與定位是電腦視覺中非常重要的一部份。通常，我們嘗試建立物體的三度空間模型（model）。這就有賴於測距器（rang finder）或立體視覺（stereopsis）為之了。但是，它們都有其缺失。因此本論文中提出數種整合影像灰度與距離資料之架構，以建立對物體之描述，並進而辨認與定位。首先，我們對所謂的融合（fusion）依其功能作一分類。〞感測器整合（sensorintegration）〞討論有關於多數個或不同種類感測器間實際操作上之合作，其目標是藉感測器間之相互導引（guidance）來提高效能，或偵測與還原感測器資料之結合。其目標是對某一種感測器資料所無法求出的。本論文中提出一架構，一攝影機與一測距器分別從不同視角去量測一物體並建立屬於自己個別之描述。經由其間相互關係（correlation） 之探討，此二種描述合而為一，其代表之物體描述將更豐富更具訊息。建立後之物體描述可以一相關式圖形（relational graph）來表示。如此，物體辨認即形成一最佳圖形化對問題（optimal graph monomorphism）。我們採用一種有效的解譯樹（interpretation tree ）搜尋法則來做辨認比對。此外，由於本系統具三度空間測量能力，因此，經由模型與實物間的轉換關係，可對物體之各特徵點（如頂角點）作精確定位，以利於機器自動化操作（robotic manipulation）。