Scholarlink

<p>Deep learning has emerged as a transformative tool for intelligent road infrastructure management, overcoming the inefficiencies of traditional manual inspection, which is often hazardous, labor-intensive, and time-consuming. This study presents a novel, real-time monitoring framework that integrates YOLOv11 for object detection, CNN-BiGRU for temporal severity prediction, and a DT-driven simulation environment to model infrastructure context. The proposed hybrid system targets critical road conditions such as potholes, surface cracks, obscured road markings, and snow-covered surfaces. Leveraging the spatial precision of YOLOv11 and the temporal consistency of CNN-BiGRU, embedded within a DT environment, the system ensures adaptive and data-driven infrastructure analysis. Evaluations were conducted on the publicly available LiRA-CD dataset, containing over 30,000 instances across diverse environmental and structural conditions. The dataset was partitioned using a 70:15:15 training-validation-test split, stratified to ensure class balance. The model was benchmarked on an Intel i7-12700K CPU, NVIDIA RTX 3090 GPU, and 32 GB DDR5 RAM. With input resolution of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(640 \times 640\)</EquationSource> </InlineEquation>, batch size of 16, 8-bit quantization, and TensorRT acceleration, the system achieved a low per-frame latency of 9.5 ms and supported real-time inference at 105 FPS. Quantitative results show the model attained a mean Average Precision (mAP@0.5) of 96.92%, mAP@[0.5:0.95] of 90.74%, and AUROC of 0.942. Additional metrics include per-class precision of 96.50%, recall of 95.29%, specificity of 96.39%, and F1-score of 96.19%. For regression-based severity prediction, the framework achieved an <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> of 0.77, with low Absolute Average Error (AAE = 0.31) and Average Squared Error (ASE = 0.34), indicating strong model generalizability and robustness. This integration of DT, YOLOv11, and CNN-BiGRU constitutes a scalable and efficient solution for proactive and real-time road infrastructure monitoring, setting a new benchmark for smart transportation systems.</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

An intelligent YOLO and CNN-BiGRU framework for road infrastructure based anomaly assessment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月21日

查看全文
  • Ainur Zhumadillayeva,
  • Tariq Ahamed Ahanger,
  • Bakhyt Matkarimov

摘要

Deep learning has emerged as a transformative tool for intelligent road infrastructure management, overcoming the inefficiencies of traditional manual inspection, which is often hazardous, labor-intensive, and time-consuming. This study presents a novel, real-time monitoring framework that integrates YOLOv11 for object detection, CNN-BiGRU for temporal severity prediction, and a DT-driven simulation environment to model infrastructure context. The proposed hybrid system targets critical road conditions such as potholes, surface cracks, obscured road markings, and snow-covered surfaces. Leveraging the spatial precision of YOLOv11 and the temporal consistency of CNN-BiGRU, embedded within a DT environment, the system ensures adaptive and data-driven infrastructure analysis. Evaluations were conducted on the publicly available LiRA-CD dataset, containing over 30,000 instances across diverse environmental and structural conditions. The dataset was partitioned using a 70:15:15 training-validation-test split, stratified to ensure class balance. The model was benchmarked on an Intel i7-12700K CPU, NVIDIA RTX 3090 GPU, and 32 GB DDR5 RAM. With input resolution of \(640 \times 640\) , batch size of 16, 8-bit quantization, and TensorRT acceleration, the system achieved a low per-frame latency of 9.5 ms and supported real-time inference at 105 FPS. Quantitative results show the model attained a mean Average Precision (mAP@0.5) of 96.92%, mAP@[0.5:0.95] of 90.74%, and AUROC of 0.942. Additional metrics include per-class precision of 96.50%, recall of 95.29%, specificity of 96.39%, and F1-score of 96.19%. For regression-based severity prediction, the framework achieved an \(R^2\) of 0.77, with low Absolute Average Error (AAE = 0.31) and Average Squared Error (ASE = 0.34), indicating strong model generalizability and robustness. This integration of DT, YOLOv11, and CNN-BiGRU constitutes a scalable and efficient solution for proactive and real-time road infrastructure monitoring, setting a new benchmark for smart transportation systems.