標線表面玻璃珠對反光性能之影響

Abstract

標線在交通工程中扮演極為關鍵的角色，主要負責引導用路人正確判斷行車方向，特別是在視野不佳的夜間或雨天環境下，反光性能更直接影響駕駛者對標線的辨識能力，同時也影響自駕車系統中機器視覺對道路資訊的準確解讀。為確保標線具備良好的可視性，需仰賴有效且即時的反光性能檢測。然而，目前常用的標線反光性能檢測方式，主要依賴反光儀等專業設備進行量測，雖然具有高準確度與穩定性，但其設備成本昂貴，限制了其在日常維護中的普及性。 本研究提出一種結合影像辨識與機器學習的反光性能預測方法。首先，透過手機拍攝標線表面之近距離影像，並使用 YOLOv8 模型進行訓練，以辨識出影像中每顆玻璃珠的位置與裸露直徑，進一步擷取影像中所有玻璃珠數量、大小與表面面積等物理特徵。為克服影像中玻璃珠尺寸無法直接得知之問題，本研究進一步建立裸露直徑對原始玻璃珠尺寸的分類模型，使用多元邏輯斯回歸將玻璃珠分至對應粒徑區間，藉此推估其原尺寸並計算沉降率。此外，為分析玻璃珠特徵與標線回歸反射輝度係數( RL )之關係，本研究分別針對標線試片與實地道路標線收集反光數據，並搭配隨機森林回歸模型建構預測模型，以驗證所提方法之準確性與穩定性。結果顯示，在 RL 預測中，標線試片資料之測試集 R² 為 0.779，實地標線資料則達到 R² 為 0.876，顯示本方法在不同場景皆能有效反映玻璃珠對反光性能之影響，並具備良好的預測能力。

Road markings play a critical role in traffic engineering by guiding road users in correctly determining driving directions. This function becomes especially crucial under low-visibility conditions such as nighttime or rainy weather, where retroreflective performance directly affects both drivers’ ability to perceive road markings and the accuracy of machine vision systems used in autonomous vehicles. To ensure sufficient visibility of road markings, effective and real-time retroreflectivity measurement is required. However, current retroreflectivity assessments rely heavily on specialized instruments such as retroreflectometers. Although these devices offer high accuracy and stability, their high cost limits their widespread adoption in routine maintenance practices. This study proposes a retroreflectivity prediction method that integrates image recognition and machine learning. Images of road marking were captured using a smartphone, and a YOLOv8 model was trained to detect each glass bead's position and exposed diameter. The model extracted every physical feature in images such as the number, size, and surface area of the beads. To address the challenge that the actual size of the beads cannot be directly observed in the image, a classification model was developed to estimate the original bead size based on exposed diameter. A multinomial logistic regression model was used to assign each bead to its corresponding size range, enabling the calculation of embedment depth. Furthermore, to analyze the relationship between bead characteristics and retroreflectivity (RL), retroreflection data were collected from both laboratory test specimens and actual road markings. A random forest regression model was employed to construct and validate the predictive framework. Experimental results show that for RL prediction, the test R² of laboratory specimens reached 0.779, while that of real-world road markings achieved 0.876. These results demonstrate that the proposed method effectively reflects the influence of glass beads on retroreflectivity across different scenarios and possesses strong predictive capability.