以影像辨識輔助軌道巡查作業──臺灣鐵路宜蘭段及臺中段實證研究

USING IMAGE RECOGNITION TO ASSIST TRACK INSPECTION OPERATIONS–EMPIRICAL RESEARCH ON YILAN SECTION AND TAICHUNG SECTION OF TAIWAN RAILWAY

本研究透過實證研究方法，探討人工智慧影像辨識方法是否有助於軌道巡查作業。國營臺灣鐵路股份有限公司（以下簡稱臺鐵公司，前身為交通部臺灣鐵路管理局）軌道巡查作業以人力目視檢查為主，容易受到行進速度、光線角度與視線角度影響，造成遺漏或誤判，且因執行效率無法提升，以致日常巡查進度有限，因此亟需發展一套自動化巡檢輔助系統，輔助巡查工作。本研究透過文獻回顧、系統設計、資料蒐集、模型訓練與實地驗證等流程，研發軌道構件缺失辨識系統，最終模型之訓練效果mAP可達96.2%。經過多次於臺鐵公司營運與養護之軌道環境進行資料蒐集與系統驗證，已確認本系統於夜間環境且行駛速度達60km/hr情況下，具有一定的檢出率與準確度。本研究結果顯示，人工智慧影像辨識方法確實有助於軌道巡檢作業，亦可做為輔助軌道巡查作業之發展基礎。

This study uses empirical research methods to explore whether artificial intelligence (AI) image recognition methods are helpful for track inspection operations. The Taiwan Railways Administration (TRA) track inspection operation is mainly based on human visual inspection, which is easily affected by the speed, the angle of light and the angle of sight, resulting in omissions or misjudgments. The inspection progress is limited, so it is urgent to develop an automatic inspection auxiliary system to assist the inspection work. In this study, through the process of literature review, system design, data collection, model training and field verification, the missing identification system of track components was developed. The mAP of the final model training effect can reach 96.2%. After repeated data collection and system verification in the track environment of the TRA's operation and maintenance, it has been confirmed that the system has a certain detection rate and accuracy in the night environment and the driving speed reaches 60km/hr. The results of this study show that the AI image recognition method is indeed helpful for track inspection operations, and can also be used as the basis for the development of auxiliary track inspection operations.