立體影像對自動特徵點提取進行三維重建

3D reconstruction with automatic feature extraction from stereo images

施宣宇(Hsuan-yu Shih)

指導教授 ： 任玄

國立中央大學/地球科學學院/太空科學研究所/碩士(2015年)

摘要

利用立體影像對重建三維目標已被研究數十年。在這期間，針對三維目標重建結構簡單且平滑的目標物已相當的成功，但對複雜與獨特的目標物，例如大範圍的地形、房子，或者是比較複雜的人臉，在三維重建仍有相當的困難。在本研究中，利用地形影像立體對與人臉影像立體對進行實驗，我們得到立體影像對後進行自動化特徵提取且進而重建三維影像，而實驗中會利用手動特徵點來輔助與確認自動特徵點的可靠性。為了達成高穩定性與效率的匹配特徵點，我們選用了尺度不變特徵轉換演算法，其利用高斯金字塔偵測特徵點，再者提供特徵描述運算子使特徵點更獨特，提高最後利用特徵匹配的準確性，最後特徵匹配依據最小歐氏距離，且為了使特徵點對更符合立體視覺的條件，加入了角度和距離的限制提高影像匹配的成功率。得到特徵匹配點後，利用雙眼立體視覺定理的公式求得三維座標進而線性內插求得三維影像，得到初始三維影像後在利用中位數濾波器去除雜訊得到較平滑的影像。在研究地形影像時，將兩種提取特徵的方法與數值地形模型進行比較，得出自動化提取特徵點的方法比手動更好、更有效率且能提取出更多的特徵點使影像更平滑，並可以利用手動特徵點的結果驗證自動化的可靠性。在人臉影像方面，發現上下垂直拍照的結果會比左右平行拍好，因為可以避免左右臉頰特徵點的誤差。

關鍵字

雙眼立體視覺 ； 尺度不變特徵轉換 ； 特徵提取 ； 三維重建

並列摘要

To construct 3-D objects from stereo images has been studied for decades. It shows some success for construction of simple objects with smooth surfaces, but in complex and unique object, such as a large terrain, house, or more complicate human face, that still remain an issue for 3-D reconstruction. In this study, we acquire the terrain aerial photos and facial images for our research, to construct 3-D model with automatic feature extraction from stereo photos without any models. To achieve robustness and efficient matching of features, we adopt Scale-Invariant Feature Transform (SIFT) algorithm to detect the feature with the Gaussian pyramid and find the feature descriptors which make the features more distinctive, hence, the feature matching can be efficiently accomplished based on Euclidean distance. Moreover, those matching pairs must satisfy the Stereoscopic vision between two images, therefore, the angle and distance constraints for features are added before feature matching to improve the accuracy of matching pairs and stereo vision. Then, the 3-D model can be reconstructed by solving the binocular stereo vision theorem from feature pairs in stereo photos, the error estimated of depth can be smooth by the median filter. Finally, we compare two feature extraction methods to the Digital Terrain Model. Compare to manual extraction, the automatic feature extracted method is much better, more efficient and provides more matched features, and therefore, it shows better results for the studies of the topographic images in 3-D construction. In accuracy assessment, we compare to the results of manual extraction to verify the reliability of the automatic method. Our experimental results in the facial images show that taking stereo pictures vertically is much better than horizontally, because vertical direction can avoid errors of feature matching in the edges of cheeks.

並列關鍵字

Binocular stereo vision ； Scale-Invariant Feature Transform (SIFT) ； feature extraction ； 3-D construction

參考文獻

[1] Quigley, N. and Lyne, J. E. Development of a Three-Dimensional Printed, Liquid-Cooled Nozzle for a Hybrid Rocket Motor. Journal of Propulsion and Power, Vol. 30, No. 6, pp. 1726-1727, 2014.

[5] Lowe, D.G. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60, 2, pp. 91-110, 2004.

[6] Lindeberg, T. 1994. Scale-space theory: A basic tool for analyzing structures at different scales. Journal of Applied Statistics, 21(2):224–270.

[11] MATLAB/C implementation of the SIFT detector and descriptor website:

[12] Mark, N. and Alberto, A. Feature extraction and image processing. Nixon, Mark S. Amsterdam; Boston: Academic, 2008.