利用深度資訊對複雜場景中的三維物體進行切割與辨識

本論文提出一個快速且有效的三維物體辨識系統，以辨識複雜場景下的三維物體。此系統可解決深度資訊(range data)中，因為測量誤差所產生的雜訊，同時提高物體被遮蔽時的辨識率，未來可將此系統應用在機器人視覺上，來進行辨識與引導動作。首先，本論文結合了適應性中值濾波器(adaptive median filter)與移動式最小平方法(moving least square)來修復因測量誤差產生的三維雜訊，以獲得正確的物體部分表面，並且提出了一個多重臨界值演算法(multilevel thresholding)，使得複雜場景變成多個單一場景，再利用深度影像中像素的連通性，來分離每個單一場景中的物體，以作為辨識之輸入。其次，為了使得目標物體在遮蔽環境下，也可以有效地被辨識出來，本論文使用了邊緣圖像(edge map)的概念，使單一物體依照其表面變化，被分割成許多不同的封閉區塊，以提高目標物體被遮蔽時的辨識率。首先，利用Canny 邊緣偵測器(Canny edge detector)來偵測出深度影像中物體的步階邊緣，然後本論文提出了計算物體表面法向量變化以形成梯度影像，再偵測出物體的屋脊邊緣(roof edge)，就形成了邊緣影像(edge image)，最後對邊緣影像使用形態學運算(morphological operator)，使邊緣影像變成邊緣圖像。然後，對每個物體的邊緣圖像中的每個封閉區塊使用區域成長法(region growing)來抽取出該區塊的特徵後，並使用多維直方圖(multidimensional histogram)統計整體特徵與區塊特徵，形成了整體直方圖(unity histogram)與部分直方圖(partial histogram)；其中，在本論文中，使用曲率之形狀指標(shape index)及法向量分量之夾角，這兩個區域特徵來表示三維物體部分表面之特徵。最後使用 -divergence計算直方圖相異程度，並且結合了幾個常用的直方圖比對方法，以計算部分直方圖的相異程度，同時提出了兩階段的辨識系統來縮短辨識所需的時間。

關鍵字

三維物體辨識；深度資訊；深度影像切割；邊緣圖像；區域特徵；多維度直方圖

並列摘要

In this thesis, a highly efficient 3D view-based object recognition system, which is to recognize 3D objects in cluttered scenes, is proposed. This system can handle the 3D noise in the range data because of measure error margin in the range finder, and increase the recognition accuracy when object is covered in cluttered scene. In the future, I hope that this recognition system will apply to the robotic vision. First of all, in order to handle the 3D noise in the range data, an algorithm which combines adaptive median filter and moving least square (MLS) is proposed in the beginning of the recognition system. After that, a multilevel thresholding method is proposed which segments a cluttered scene into several monotonous scenes, and then separates each object in the scene by using the 8-connected component of pixels in range image. These objects will be the input of the recognition system. More importantly, in order to recognize objects which are covered in cluttered scene, a concept of edge map is applied in this thesis. Then, extract the feature belongs to each closed region in the edge map by using region growing method, and calculate the features to create unity histogram and partial histogram by using multidimensional histogram; moreover, the local feature is presented as features of 3D object’s surface; however, in order to increase the speed during the recognition, a two-step recognition system is presented in this thesis.

並列關鍵字

3Ｄ object recognition ； range data ； range image segmentation ； edge map ； local feature ； multidimensional histogram

參考文獻

[1] O. Carmichael, D. Huber, and M. Hebert, “Large data sets and confusing scenes in 3-D surface matching and recognition,” Int. Conf.3-D Digital Imaging and Modeling, pp. 358-367, Ottawa, Ont., Canada, Oct. 1999.

[2] G. C. Sharp, S. W. Lee, and D. K. Wehe, “Maximum-likelihood registration of range images with missing data,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 30, no. 1, pp. 120-130, Jan. 2008.

[3] H. Gharavi and S. Gao, “3-D motion estimation using range data,” IEEE Trans. Intelligent Transportation Systems, vol. 8, no. 1, pp. 133-143, March 2007.

[4] S. Hussmann and T. Liepert, “Three-dimensional TOF robot vision system,” IEEE Trans. Instrumentation and Measurement, vol. 58, pp.141-146, Jan. 2009.

[7] R. L. Hoffman and A. K. Jain, “Segmentation and classification of range images,” IEEE Tran. Pattern Analysis and Machine Intelligence, vol. 9, no. 5, pp. 608-620, Sept. 1987.

國際替代計量

利用深度資訊對複雜場景中的三維物體進行切割與辨識

全文下載

主題瀏覽