Scholarlink

<p>In Advanced Driver Assistance Systems (ADAS) and autonomous vehicles, accurate interpretation of road markings such as stop lines, lane boundaries, and pedestrian crossings is essential for maintaining navigation precision and traffic safety. These systems rely on the consistent detection of such markings to determine vehicle positioning and make real-time driving decisions. To address this critical requirement, the present study introduces a deep learning (DL) model based on the EfficientDet architecture for detecting road features under diverse environmental conditions. The model is trained on the ROAD MARK Dataset, which includes 2,892 annotated samples from 14 distinct classes captured in varied real-world settings. Unlike prior studies focused primarily on lane detection or constrained lighting conditions, this work addresses the broader challenge of detecting multiple classes of road markings such as arrows, pedestrian crossings, and lane indicators, across a wide range of environmental scenarios. The model also integrates Contrast Limited Adaptive Histogram Equalization (CLAHE) in preprocessing and employs stratified class-balanced training, improving robustness in low-visibility and class-imbalanced conditions. EfficientDet combines the EfficientNet backbone for effective feature extraction with a Bi-directional Feature Pyramid Network (BiFPN) for multi-scale feature fusion, enabling it to handle complex and variable marking patterns. The system outputs both class labels and bounding boxes for detected road features. The model achieved outstanding performance across multiple evaluation metrics, including 98.90% accuracy, 98.13% precision, 97.95% recall, 98.10% F1 score, an Intersection over Union (IoU) of 0.85, and a mean Average Precision (mAP) of 0.928. A five-fold cross-validation confirms the model’s statistical robustness, reporting narrow confidence intervals (e.g., ± 0.07 for accuracy), while external validation on the Tsinghua-Tencent 100&#xa0;K dataset affirms generalization capabilities. These results affirm the model’s effectiveness in delivering reliable and accurate road marking recognition for real-time autonomous driving applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DeepVisionMark: Enhancing Autonomous Navigation Through Road Mark Detection Using Advanced Deep Learning Model

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • Sruthy S,
  • Anju J Prakash

摘要

In Advanced Driver Assistance Systems (ADAS) and autonomous vehicles, accurate interpretation of road markings such as stop lines, lane boundaries, and pedestrian crossings is essential for maintaining navigation precision and traffic safety. These systems rely on the consistent detection of such markings to determine vehicle positioning and make real-time driving decisions. To address this critical requirement, the present study introduces a deep learning (DL) model based on the EfficientDet architecture for detecting road features under diverse environmental conditions. The model is trained on the ROAD MARK Dataset, which includes 2,892 annotated samples from 14 distinct classes captured in varied real-world settings. Unlike prior studies focused primarily on lane detection or constrained lighting conditions, this work addresses the broader challenge of detecting multiple classes of road markings such as arrows, pedestrian crossings, and lane indicators, across a wide range of environmental scenarios. The model also integrates Contrast Limited Adaptive Histogram Equalization (CLAHE) in preprocessing and employs stratified class-balanced training, improving robustness in low-visibility and class-imbalanced conditions. EfficientDet combines the EfficientNet backbone for effective feature extraction with a Bi-directional Feature Pyramid Network (BiFPN) for multi-scale feature fusion, enabling it to handle complex and variable marking patterns. The system outputs both class labels and bounding boxes for detected road features. The model achieved outstanding performance across multiple evaluation metrics, including 98.90% accuracy, 98.13% precision, 97.95% recall, 98.10% F1 score, an Intersection over Union (IoU) of 0.85, and a mean Average Precision (mAP) of 0.928. A five-fold cross-validation confirms the model’s statistical robustness, reporting narrow confidence intervals (e.g., ± 0.07 for accuracy), while external validation on the Tsinghua-Tencent 100 K dataset affirms generalization capabilities. These results affirm the model’s effectiveness in delivering reliable and accurate road marking recognition for real-time autonomous driving applications.