請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/89852
標題: | 架設於火車頭之單目視覺鐵道障礙物測距系統 Monocular Railway Obstacle Ranging System on Locomotives |
作者: | 李承翰 Cheng-Han Li |
指導教授: | 莊裕澤 Yuh-Jzer Joung |
關鍵字: | 單目視覺,鐵路障礙物偵測,影像測距,物件辨識, Monocular Vision,Railway Obstacles Detection,Vision-Based Distance Estimation,Object Detection, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 台鐵因長時間行駛在開放、非高架的環境,常常受到外物侵入的困擾,造成輕則誤點、重則人員傷亡的事故。然而,台鐵非平交道段現行的障礙物偵測方式主要依靠人員發現後撥打緊急專線,耗時且仰賴人力。駕駛員往往會是最先觀測到障礙物的人員,故障礙物偵測系統應架設在火車頭,輔助駕駛員識別障礙物,並提供有用的資訊協助其制定避障決策。
因此,本研究設計了一套易於部署、架設於火車頭的障礙物辨識與測距系統,可協助駕駛員辨識火車前方的障礙物,其種類以及距離,有了這些資訊即可推知火車撞上障礙物的時間,駕駛員可依此制定出最適當的避障策略。 本研究的系統以成像原理蘊含的幾何關係,利用軌距和物件框作為參考物進行距離估計,其中軌距以Canny Edge Operator和Hough transform提取出的鐵軌計算得出,物件框則以當今效能與精準度兼具的YOLOv7物件辨識演算法提取。我們利用實地錄製的影像作為資料集進行實驗,接著就「滯空物判斷與距離估計」、「以物件尺寸提升距離估計準度」進行延伸探討,最終從實驗結果中,我們得出本系統在距離估計準確度方面有良好的表現,與其他類似研究相比,還可兼備物件辨識與滯空物判斷的功能。 Trains of Taiwan Railway run in an open, non-elevated environment, so it is sometimes disturbed by the intrusion of foreign objects, which could cause delays or even accidents with injuries or fatalities. However, the current method of detecting obstacles on the non-level crossing sections of the Taiwan Railway relies mainly on personnel to call the emergency hotline after observing obstacles, which is time-consuming and relies on labor. Drivers are often the first to observe obstacles, therefore, obstacle detection system should be installed at the front of the train to assist drivers in identifying obstacles and provide useful information to help them make decisions on obstacle avoidance. In this study, an easily deployable obstacle recognition and distance measurement system is designed to help drivers recognize the obstacles in front of the train, their types and distances. With these information, the system can infer the time when the train hits the obstacles so as to help the drivers take appropriate actions accordingly. The system in this study utilizes the geometrical relationship embedded in the imaging principle to estimate the distance using the track gauge and the bounding box of the detected objects as references, where the track gauge is calculated from the railroad tracks extracted by Canny Edge Operator and Hough transform, and we used YOLOv7 as object detection model, which is the most efficient and accurate algorithm nowadays. We utilize the field-recorded images as a data set for the experiments, and then extend the discussion on "distinguishing and ranging objects in the air" and "improving distance estimation accuracy with object size". Our experiments show that, compared with existing methods, our system has good accuracy for estimating object distances. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/89852 |
DOI: | 10.6342/NTU202303543 |
全文授權: | 未授權 |
顯示於系所單位: | 資訊管理學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 42.69 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。