使用顯著幾何圖形描述子之三維點雲物體辨識

Abstract

由於三維物體具有種類繁多的樣貌與形狀，因此在電腦視覺領域中，三維物體辨識是一個極具挑戰性的研究主題。基於物體模型本身的點雲資料辨識，牽涉到三維物體的偵測與定位。然而，許多三維物體辨識的方法，例如旋轉影像法，直接使用表面疊合或是藉由尋找對應點的方式，進行物體表面的比對。像這樣的辨識方法都需要進行大量的計算，特別是使用於點雲資料時。為了能更準確且有效率的辨識點雲資料，本論文提出了一個顯著幾何描述子，加上一個具有三階段架構的辨識系統，來解決這個問題。在描述階段，經由多尺度重要點選取、法向量角度變化以及幾何圖形建立，可以將物體表面的特徵擷取出來。在偵測階段，一個基於詞袋模型的表示法，搭配簡單貝氏分類器進行物體的偵測，並且能限制辨識階段在資料庫中搜尋的範圍。最後，在辨識階段中，再以對應點搜尋法驗證辨認出的物體，同時得到其估測的姿態。從實驗的結果中可看出，本論文提出的方法能精確的辨識在場景中的三維物體；在完整物體的辨識率可以高達95%，而在部分表面的辨識率則有79%。同時，本論文的方法在低解析度時較旋轉影像法更能辨識部分的物體表面。實驗中也提出了物體檢索與實際場景辨識的兩個應用例，顯示本論文提出的方法具備實用性，並且能可靠地應用在資料庫的檢索與真實點雲場景的辨識。

Object recognition is one of the most challenging research fields in computer vision, since objects have an enormous variety of appearances and shapes. Model-based approaches for 3-D object recognition involve the detection and localization of 3-D models. Moreover, many of them directly use surface alignment or point correspondence to match surfaces, for instance, the spin-image method. However, such processes of recognition need high computational costs, especially for dense point clouds models. To provide more accurate and efficient recognition, a three-phase scheme of 3-D object recognition system with a salient geometric graph descriptor is presented. In the representation phase, surface features are extracted from point clouds by using multi-scale salient point selection and differential normal calculation. A geometric graph is then constructed to describe the surface. In the detection phase, a bag-of-words representation and a naive Bayes classifier are presented to detect objects and limit the searching range in the database. Finally, the recognition phase is proposed to obtain the identified object with an estimated pose. The experimental results show that the recognition rates achieve 95% in complete object recognition and 79% in partial object recognition. Also, it performs better than spin-image method. Furthermore, two applications are demonstrated to show the practicability and the reliability in large-scale datasets and real data.