基於GPU平行計算之視覺里程計

Visual Odometry Based on GPU Computation

劉申禾(Liu Shen-Ho)

指導教授 ： 許陳鑑 ； 王偉彥

國立臺灣師範大學/科技與工程學院/電機工程學系/碩士(2018年)

https://doi.org/10.6345/THE.NTNU.DEE.006.2018.E08

摘要

本論文係針對視覺里程計(Visual Odometry, VO)系統進行改良，加入地標管理、key-frame選擇機制以及攝影機定位之修正模型來提高視覺里程計的定位準確性，並利用GPU平行運算的優勢，實現一高運算效率的系統，使機器人在行走中，能夠即時地推算自身相對於初始位置的狀態。視覺里程計的改良主要是利用SURF演算法提取特徵點，並利用比對窮舉法比對當下特徵與地標， key-frame選擇機制以避免多餘的運算量，並加入地標管理機制來濾除、新增地標點，以及判斷攝影機有效解之機制來解決P3P演算法兩面解之問題，最後使用P3P及RANSAC演算法推算出攝影機位置，為解決累積誤差的問題本論文加入攝影機定位之修正模型來提升視覺里程計的定位準確性。為了達到即時性，本論文利用GPU平行運算的優勢，執行SURF演算法並搭配比對窮舉法找出定位之特徵點、以及針對P3P以及RANSAC適合之架構進行設計，並利用異質運算，亦即CPU搭配GPU，將整個VO系統實現在TX2嵌入式系統上，因此，整體運算的效率得以大幅提升。實驗結果顯示，相較於只有使用CPU運算速度而言，異質運算在整體的效能提升了約80~90倍之多，顯示本論文基於GPU平行計算之視覺里程計可提供一個低成本、低功耗、可攜性、高效能且即時性之視覺里程計系統，達到即時視覺里程計之目的。

關鍵字

視覺里程計 ； SURF ； Perspective-Three-Points(P3P) ； GPU ； TX2

並列摘要

In this thesis, an improved visual odometry (VO) system is proposed based on novel map management, key-frame selection, and a camera pose correction model. To enhance computational efficiency of the VO system, the proposed approach implements a graphics processing unit (GPU), SURF feature detection and description algorithms to extract features from an image; an exhaustive search algorithm is introduced to match features. To minimize computation time, a key-frame selection mechanism is proposed to distinguish key-frames among the input images. Moreover, map management is proposed to filter out unstable landmarks and add features for a reliable estimation of the relative camera pose. Additionally, a method to validate the camera pose is proposed to solve the problem of two possible poses resulting from the use of the perspective-three points (P3P) algorithm. Estimation accuracy is improved by basing the enhanced VO system on a camera pose correction model. Furthermore, to accelerate execution efficiency, GPU implementation of the proposed VO system is developed, taking advantage of parallel computation. In this thesis, the entire VO system is implemented on a TX2 embedded system under a heterogeneous computing architecture to increase the efficiency of the overall system. Several experiments are conducted for validation using an ASUS Xtion 3D camera and a laptop. Average errors of pose estimations are compared with the conventional VO to show the effectiveness of the proposed VO system. By utilizing GPU for the algorithm including SURF, exhaustive search, and pose estimation, a real-time VO system is realized with the performance with low in cost and power consumption, high processing efficiency, and easier portability. Experimental results show that its required computation time for overall performance based on heterogeneous computing is approximately 80 to 90 times faster than using only CPU computing.

並列關鍵字

Visual Odometry ； SURF ； Perspective-Three-Points(P3P) ； GPU ； TX2

參考文獻

[1] C. Forster, M. Pizzoli, and D. Scaramuzza, “SVO: fast semi-direct monocular visual odometry,” IEEE International Conference on Robotics and Automation (ICRA), Hong Kong, June, 2014.

[2] D. Scaramuzza, and F. Fraundorfer, “Visual odometry: part I: the first 30 years and fundamentals,” IEEE Robotics & Automation Magazine, Vol. 18, pp. 80-92, 2011.

[3] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International Journal of Computer Vision, Vol. 60, No. 2, pp. 91-110, 2004.

[4] H. Bay, A. Ess, T. Tuytelaars, and L. V. Gool, “SURF: speeded up robust features,” European Conference on Computer Vision (ECCV), Graz, May, 2006, pp. 404-417.

[5] K. Mikolajczyk, T. Tuytelaars, C. Schmid, A. Zisserman, J. Matas, F. Schaffalitzky, T. Kadir, and L. Van Gool, “A comparison of affine region detectors,” International Journal of Computer Vision, Vol. 65, No. 1-2, pp.43-72, 2005.