Please use this identifier to cite or link to this item: http://hdl.handle.net/11455/97068
標題: 基於RCNN電腦視覺技術之傾斜角度測量系統
A Tilt Angle Measurement System using RCNN Computer Vision Technique
作者: 余修瑋
Hsiu-Wei Yu
關鍵字: 計算機視覺;更快的區域卷積神經網絡;資料融合;感測器;Computer Vision;Faster R-CNN;Data fusion;Sensor
引用: [1] W. Yang, B. Fang, Y. Y. Tang, J. Qian, X. Qin, W. Yao, 'A robust inclinometer system with accurate calibration of tilt and azimuth angles', IEEE Sensors J., vol. 13, no. 6, pp. 2313-2321, Jun. 2013. [2] J. Mercer, P. Hambling, R. Zeller, S. Ng, G. Brune, L. Moore, System for tracking and/or guiding an underground boring tool, Mar. 2000. [3] J.Matej, “Determination of forestry machine’s tilt angle using camera and image processing,” Comput. Electron. Agric., vol. 109, pp. 134–140, 2014. [4] A. Godfrey, R. Conway, D. Meagher, G. Ólaighin, 'Direct measurement of human movement by accelerometry', Med. Eng. Phys., vol. 30, no. 10, pp. 1364-1386, 2008. [5] F. Bagala, V. Fuschillo, L. Chiari, A. Cappello, 'Calibrated 2D angular kinematics by single-axis accelerometers: From inverted pendulum to N-link chain', IEEE Sensors J., vol. 12, pp. 479-486, Mar. 2012. [6] J. R.Blum, D.Greencorn, and J. R.Cooperstock, “Smartphone sensor reliability for augmented reality applications,” Mob. ubiquitous Syst. Comput. networking, Serv., pp. 127–138, 2013. [7] “The Compass Within – sciencewriter.org.” [Online]. Available: https://sciencewriter.org/the-compass-within-understanding-animals-magnetic-sense/. [8] W.Wiltschko and R.Wiltschko, “Magnetic orientation and magnetoreception in birds and other animals,” J. Comp. Physiol. A Neuroethol. Sensory, Neural, Behav. Physiol., vol. 191, no. 8, pp. 675–693, 2005. [9] “An Overview of the Earth’s Magnetic Field.” [Online]. Available: http://www.geomag.bgs.ac.uk/education/earthmag.html. [10] Alan E.Mussett and M. A.Khan, Looking into the earth: an introduction to geological geophysics. Cambridge University Press, 2000. [11] F. Y.Sung, S. H.Fang, andY. R.Chien, “An experimental study of MEMS-based magnetometers on Android mobile phones,” in Digest of Technical Papers - IEEE International Conference on Consumer Electronics, 2014, pp. 227–228. [12] R. Girshick, J. Donahue, T. Darrell, J. Malik, 'Region-based convolutional networks for accurate object detection and segmentation', TPAMI, 2015. RCNN [13] R. Girshick, 'Fast R-CNN', Proc. IEEE Int. Conf. Comput. Vis., pp. 1440-1448, 2015. [14] S. Ren, K. He, R. Girshick, J. Sun, 'Faster R-CNN: Towards real-time object detection with region proposal networks', NIPS, 2015. [15] G. A. Korn, T. M. Korn, Mathematical Handbook for Scientists and Engineers, New York:McGraw-Hill, 1968.
摘要: 
台灣是多地震的國家,建築物、路燈、工地的基樁(Pile)等容易因為地震而產生傾斜,如果傾斜到了某個程度,容易倒塌造成危險,但是如果要進行量測需要透過經緯儀(Theodolite)或鉛錘觀測垂直度,測量過程耗人力且耗時。故本論文提出一種較簡便的非接觸式測量方式,利用遠距離的相機拍攝的方向和影像的資料來目標物的傾斜角度與傾斜方位,以便進行後續的處理或維護。
首先,我們設計了一套公式,透過兩張不同角度拍攝的相片用投影的原理來計算出目標物的傾斜角度與傾斜方向。由於感測器容易受到外在磁場的干擾,故我們利用增加不同拍攝方位的方式使資料增加,以提高估測的準確度。
為了精確的抓取到複雜圖片中目標物的位置,我們採用Faster R-CNN目標檢測演算法進行目標物的辨識,再利用HSV特徵與霍夫變換(Hough Transform)將目標物在照片中的投影傾斜角計算出來,藉由公式計算出目標物的傾斜角度與傾斜方位的估算值。最後我們利用資料融合(Data fusion)的方式精練估算值,具體來說,我們利用機器學習(Machine Learning)中迴歸分析(regression analysis)的技術,分析估測值的誤差,找出提升精確度的特定角度關係,並設計一套回授強化的演算方法,使觀測者可以依據之改變觀測方位或挑選取樣資料,以快速提升估算的精確度。

Taiwan is located on the circum-Pacific seismic zone and earthquakes occur quite frequently. As a result, buildings, street lights, and piles are prone to tilt because of the earthquakes. If they tilt to a certain extent, there is a high risk of collapsing. To measure the tilt angle, devices like theodolites or posting plumbs are commonly used and the process is generally expensive, time-consuming, dangerous and even stringent in mountainous terrain. Therefore, this paper introduces a relatively simple, flexible, non-contact measurement method – using devices like a smart phone with a camera and a geomagnetic sensor to estimate the tilt angle and tilt azimuth of a target so as to automatically detect potential crisis for subsequent maintenance.
First, according to the projection theory, we derive formulas to estimate the tilt angle and tile azimuth by using two shootings at different positions. Specifically, we capture an image as well as the shooting azimuth at each shooting and use image processing technique to calculate the projected tilt angle in each photo. Then, we apply our formula to compute the estimated tilt angle and tilt azimuth. Moreover, we apply the Faster R-CNN target detection algorithm to identify the target and then use the HSV feature and Hough Transform to obtain projected tilt angle of the target in a complicated photo. Since the sensors are easy to be interfered with the external magnetic field, we derive a selection rule empirically and further design a rapid reinforcement method to fuse multiple estimated tilt angles and azimuths to improve the estimation accuracy.
URI: http://hdl.handle.net/11455/97068
Rights: 同意授權瀏覽/列印電子全文服務,2020-08-30起公開。
Appears in Collections:電機工程學系所

Files in This Item:
File Description SizeFormat Existing users please Login
nchu-106-7104064208-1.pdf3.28 MBAdobe PDFThis file is only available in the university internal network    Request a copy
Show full item record
 

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.