三維物件之辨識與姿態估測

3D Object Recognition and Pose Estimation

邱駿展(Chun-Chan Chiu)

指導教授 ： 陳金聖

國立臺北科技大學/機電學院/自動化科技研究所/碩士(2009年)

https://doi.org/10.6841/NTUT.2009.00047

摘要

物體識別(Object recognition)與姿態估測在目前影像伺服與機器視覺等相關應用中仍是一個重要且基礎的問題。本篇論文主要在發展一個三維物體識別與姿態估測之視覺系統，對特定的目標物進行識別且鎖定，並利用立體視覺量測的方式獲得該物體的三維座標與姿態。
本論文之視覺系統之架構可分為離線訓練階段與線上操作階段，離線訓練階段進行目標物的特徵提取(Feature detector)與特徵描述(Feature descriptor)計算、建立粗與細之搜尋結構和相機校正，線上操作階段進行輸入影像的特徵提取與特徵描述計算、目標物識別和估測目標物位置與姿態。因本視覺系統同時考量穩定性與即時性，所以本論文提出改良的直覺式角點外觀偵測方式，以達到快速特徵點提取的效果，特徵描述部份使用SIFT的描述方式，並透過PCA的方式來降低其描述空間的維度，不同的是，離線訓練階段使用多解析度的特徵描述範圍，但線上操作階段只使用單一解析度的特徵描述範圍，在目標物識別時本論文提出階層式的比對方式，此方式利用縮減搜尋空間來加快比對速度。由實驗結果顯示，使用本論文所提出之特徵點能夠迅速識別出單一目標物，而在一般雜亂之環境下仍然有足夠的能力識別出目標物，並能進一步求得其三維空間上的座標位置，了解該目標物的在空間上的姿態。

關鍵字

物體識別 ； 視覺追蹤 ； 特徵比對

並列摘要

3D object recognition and pose estimation is a fundamental technique for many applications of machine vision, including target tracking, visual servo, robot vision, just to name a few. This thesis proposes a high speed and robust vision system for 3D objects recognition and poses estimation based on stereo projective.
The algorithm architecture of this paper consists of two phases: (1) the off-line training phase, and (2) the on-line operating phase. In the training phase, the data set corresponding to the targets are collected, then the feature points are detected and the feature descriptions are represents. With this information, the algorithm creates the hierarchical structures of feature descriptions to improve the speed of patch matching. In the other hand, the camera calibration is also finished in this phase. For the operating phase, the target images are input and the same processes for feature detection and representation are done. Finally, the recognition algorithm based on the hierarchical structures improves the performance of object detection and the 3D pose is further estimated. This thesis proposes the modified intuitive corner detection to quickly extract the features, and the feature descriptions based on SIFT and PCA are applied. In patch matching process, the multi-resolution patch in training phase and single resolution patch in operating phase are loaded respectively and the two stages matching, coarse and fine matching based on hierarchical structures of feature descriptions to reduce the range of candidates, are proposed to reduce the matching computation time. Experimental results show that the proposed algorithms can rapidly detect the target even in the complex environment. Furthermore, the pose of 3D object can be easily estimated using the transformation formula from 2D to 3D.

並列關鍵字

object recognition ； visual tracking ； feature matching

參考文獻

[3] M. Pressigout and E. Marchand, "Real-time hybrid tracking using edge and texture information, " International Journal of Robotics Research, Vol. 26, No. 7, pp.689-713, July 2007.

[4] T. Drummond and R. Cipolla, "Real-Time visual tracking of complex structures," IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 24, No. 7, pp.932-946, July 2002.

[5] D. Kragic, M. Bjorkman, H.I. Christensen, J. Eklundh, "Vision for robotic object manipulation in domestic settings," Autonomous Systems 52, pp.85-100, June 2005.

[7] J. Wang, H. Zha and R. Cipolla, "Coarse-to-fine vision-based localization by indexing scale-invariant features," IEEE Transactions on Systems, Part B: Cybernetics, Vol. 36, No. 2, April 2006.

[8] S. Ahn, M. Choi, J. Choi and W.K. Chung, "Data association using visual object recognition for EKF-SLAM in home environment," Proc. IEEE/RSJ International Conference on Intelligent Robots and Syatems, pp.2588-2594, October 2006.