<rss version="2.0" xmlns:atom="https://www.w3.org/2005/Atom">
  <channel>
    <title>Transport Research International Documentation (TRID)</title>
    <link>https://trid.trb.org/</link>
    <atom:link href="https://trid.trb.org/Record/RSS?s=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" rel="self" type="application/rss+xml" />
    <description></description>
    <language>en-us</language>
    <copyright>Copyright © 2026. National Academy of Sciences. All rights reserved.</copyright>
    <docs>http://blogs.law.harvard.edu/tech/rss</docs>
    <managingEditor>tris-trb@nas.edu (Bill McLeod)</managingEditor>
    <webMaster>tris-trb@nas.edu (Bill McLeod)</webMaster>
    <image>
      <title>Transport Research International Documentation (TRID)</title>
      <url>https://trid.trb.org/Images/PageHeader-wTitle.jpg</url>
      <link>https://trid.trb.org/</link>
    </image>
    <item>
      <title>SIT-GAN: A Slice Image Translation Network for Asphalt Mixture</title>
      <link>https://trid.trb.org/View/2685433</link>
      <description><![CDATA[In asphalt mixture research, deep learning networks have been widely applied to analyze slice images, serving as a foundation for numerous further valuable research. However, the effectiveness of these networks is hindered by the high cost of acquiring large-scale datasets, which is a great challenge for slice image dataset construction. To address this challenge, we propose Slice Image Translation Generative Adversarial Network (SIT-GAN), a novel slice image translation network designed to generate realistic images across various mixture types. Leveraging an innovative hierarchical feature extraction module (HFEM), SIT-GAN effectively captures the complicated multiscale features. Both qualitative and quantitative evaluations, as well as validation experiments, demonstrate the superior generation quality of SIT-GAN. Moreover, we show that using SIT-GAN for data augmentation significantly enhances the performance of classification and segmentation networks, providing a robust foundation for intelligent analysis in pavement engineering.]]></description>
      <pubDate>Thu, 30 Apr 2026 09:11:44 GMT</pubDate>
      <guid>https://trid.trb.org/View/2685433</guid>
    </item>
    <item>
      <title>YOLOv11-Based Enhanced Framework for Robust Blasthole Detection in Intelligent Tunnel Environments</title>
      <link>https://trid.trb.org/View/2684183</link>
      <description><![CDATA[Intelligent tunneling has emerged as a critical frontier in modern infrastructure engineering, where automation is essential to improving construction efficiency, precision, and safety. Within the widely adopted drill-and-blast method, accurate blasthole detection remains a major challenge due to harsh tunnel conditions such as dust interference, irregular rock textures, and low-resolution imaging. To address these issues, we propose YOLO-BD, a deep learning–based detection framework that extends YOLOv11 with three targeted architectural enhancements: the SPD_Conv module for information-preserving down-sampling, the C3K2_PPA attention mechanism for multiscale feature refinement, and the MB_Conv module for lightweight yet expressive feature representation. In addition, an improved WIoU_v3 loss function is introduced to enhance localization robustness under noisy and complex environmental conditions. Experimental evaluations on a custom tunnel blasthole data set show that YOLO-BD achieves 94.14% precision, 82.12% recall, 87.28% mAP50, and 49.31% mAP50:95, outperforming its YOLOv11 backbone by 4.35% and 2.97% on the respective mAP metrics. Visualization analyses further confirm YOLO-BD’s superior localization accuracy and reduced false detections under degraded imaging conditions. Comprehensive experiments on public benchmarks, including PASCAL VOC and RSOD, validate the model’s strong generalization capability, with YOLO-BD consistently surpassing baseline models across all detection metrics. Ablation studies and Grad-CAM visualizations substantiate the effectiveness of each proposed module. Overall, YOLO-BD offers a robust, accurate, and deployable solution for real-time blasthole detection in intelligent tunneling systems, with strong potential for integration into autonomous robotics and broader industrial inspection applications.]]></description>
      <pubDate>Thu, 30 Apr 2026 09:11:44 GMT</pubDate>
      <guid>https://trid.trb.org/View/2684183</guid>
    </item>
    <item>
      <title>An Efficient and Robust Model for Precise Surface Crack Segmentation on Concrete Bridge Structures</title>
      <link>https://trid.trb.org/View/2683142</link>
      <description><![CDATA[This paper proposes an efficient and robust model named bridge crack segmentation network (BCS-Net), tailored for high-precision segmentation of surface cracks on concrete bridges. BCS-Net adopts an encoder-decoder framework augmented by a multiscale feature extractor. A novel learnable hybrid sampler (LHS) is introduced to adaptively coordinate two complementary sampling strategies, reducing information loss from spatial compression and expansion while enhancing the preservation and reconstruction of crack details. Additionally, a composite enhancement module (CEM) is designed to strengthen the model’s focus on crack-relevant regions by synergistically integrating channel-wise and spatial attention mechanisms, effectively retaining critical structural cues while suppressing irrelevant features. Extensive experiments on both self-built and public data sets demonstrate that the proposed BCS-Net achieves an optimal trade-off between segmentation accuracy and computational efficiency, surpassing several mainstream segmentation models (U-Net, DeepLabv3+, SegFormer, HRNet-OCR, and Swin-Unet) and crack-specific models (CDU-Net and DTrC-Net). Specifically, BCS-Net attains F-measure scores of 89.53%, 86.31%, and 73.55%, and intersection-over-union (IOU) scores of 81.04%, 75.92%, and 58.16% on data sets comprising 1,000 self-built samples, 700 public samples, and 1124 CRACK500 samples, respectively. Moreover, BCS-Net delivers an inference speed of 13.84 frames per second (FPS), corresponding to a processing time of approximately 72 milliseconds per 1,024×512×1 resolution image, highlighting its strong potential for real-time deployment in automated bridge inspection scenarios.]]></description>
      <pubDate>Thu, 30 Apr 2026 09:11:44 GMT</pubDate>
      <guid>https://trid.trb.org/View/2683142</guid>
    </item>
    <item>
      <title>Connected Car-Based Pavement Roughness Prediction Using CNN Model and Signal Decomposition Technique</title>
      <link>https://trid.trb.org/View/2592034</link>
      <description><![CDATA[Accurate pavement roughness assessment is critical for effective pavement management. Connected car data provides a cost-effective way to predict the pavement roughness based on the acceleration data. This paper proposes a novel prediction model integrating the variational mode decomposition (VMD) with convolutional neural network to predict road profile. A particle swarm optimization algorithm is employed to determine the optimal parameters of VMD. The decomposition is performed for the acceleration signal and its double integration item. The model is validated using the field measurement data. Its prediction performance is compared with that of artificial neural networks and a convolutional neural network (CNN) model utilizing discrete wavelet transformation. Furthermore, the predicted profile is validated by the ground truth values from both the spatial and frequency domains, with the international roughness index also being calculated. The results indicate that the proposed CNN model provides a promising estimation of the road profile across different roughness levels.]]></description>
      <pubDate>Tue, 10 Mar 2026 09:57:53 GMT</pubDate>
      <guid>https://trid.trb.org/View/2592034</guid>
    </item>
    <item>
      <title>Pavement Crack Segmentation Based on Synthetic Data Sets and Unsupervised Domain Adaptation</title>
      <link>https://trid.trb.org/View/2592022</link>
      <description><![CDATA[This study proposes a novel pavement crack segmentation methodology that integrates synthetic data sets with unsupervised domain adaptation to address annotation challenges in pavement crack data sets and enhance segmentation performance. An automated program has been developed to generate synthetic pavement crack images along with high-precision semantic segmentation labels at a speed of 0.5 s per image. These synthetic images accurately replicate the irregular geometry, randomness, and variability of real pavement cracks. The CycleGAN unsupervised domain adaptation network, which incorporates the Structural Similarity Index Measure (SSIM) in its loss function, ensures that the translated images from real crack images preserve the structural characteristics of real images while enhancing domain adaptability. Applying a segmentation network to the translated images effectively mirrors segmenting real images. Segmentation performance across six semantic segmentation networks—U-net, Erfnet, DeepLabV3, FCN, PSPNet, and SegFormer—trained with synthetic images shows significant improvements when applied to the translated images, compared with directly segmenting real images, with average increases of 16.55% in mFscore and 12.62% in mIoU. Moreover, 10 domain adaptive segmentation networks—Daformer, Hrda, IAST, CBST, PyCDA, TransNorm, CLAN, FADA, AdaptSeg, and DDC—are evaluated to assess the performance of the synthetic data set. Among these, Daformer exhibits exceptional performance with an mFscore of 79.52% and an mIoU of 69.67%. In addition to the CRACK500 data set, the generalizability of the synthetic data set is validated using two other real pavement crack data sets: AEL and GAPs384. An ablation study on the CycleGAN network, trained both with and without the synthetic data set, demonstrates that incorporating the synthetic data set significantly enhances the quality of the translated images, thereby improving the model’s segmentation performance on real crack data sets. These findings highlight the significant potential of synthetic data sets in semantic segmentation tasks.]]></description>
      <pubDate>Tue, 10 Mar 2026 09:57:53 GMT</pubDate>
      <guid>https://trid.trb.org/View/2592022</guid>
    </item>
    <item>
      <title>Causality Constrained Deep Learning for Explainable Risk Analysis of Bridge Pier Settlement</title>
      <link>https://trid.trb.org/View/2632890</link>
      <description><![CDATA[Deep learning (DL) has significantly enhanced risk management in highway bridge operation. Although DL models exhibit strong predictive capabilities, their outputs derive from “black box” processes and cannot characterize uncertainty, leading to an unreliable risk decision-making. Thus, this study explores the following question: How to accurately predict bridge pier settlement in a trustworthy approach, and quantify risk uncertainty? To address this, we develop a causality constrained gated recurrent unit model to capture causal relationships inherent in monitoring data, further collaborated with a sampling method to analyze pier settlement risk under uncertainty. The proposed approach encompasses (1) introducing a transfer entropy (TE)-based causal inference method to quantify the causal relationships between factors of highway bridge operation system; (2) developing a TE-constrained loss function to guide model gradient updates to be consistent with causalities; and (3) applying the Markov chain Monte Carlo sampling to assess pier settlement risk considering uncertainty. Two real-world bridge projects in Guizhou, China, were used to validate the proposed method. The results indicate that the proposed approach achieves high accuracy in predicting pier settlement and quantifies its risk uncertainty effectively.]]></description>
      <pubDate>Mon, 02 Feb 2026 09:32:47 GMT</pubDate>
      <guid>https://trid.trb.org/View/2632890</guid>
    </item>
    <item>
      <title>TailorAlert: Large Language Model–Based Personalized Alert Generation System for Road Infrastructure Management with Digital Twins</title>
      <link>https://trid.trb.org/View/2632716</link>
      <description><![CDATA[Digital twins (DTs) have emerged as a promising tool in road infrastructure management, enabling real-time monitoring, proactive maintenance, and data-driven decision making. However, existing digital twin systems primarily rely on generic, nonpersonalized alerts to proactively deliver information to users, failing to differentiate between the diverse information needs of different stakeholders. This results in information overload for some users and insufficient detail for others, leading to delayed response times and inefficiencies in operations and maintenance (O&M) workflows. This paper presents a proof-of-concept study on a large language model-based personalized alert generation system, TailorAlert, which integrates with digital twins to provide role-specific alerts. The system utilizes a large language model (LLM) to transform general digital twin alerts into role-specific messages, ensuring that each stakeholder receives the right level of detail tailored to their responsibilities and skill levels. The system is evaluated across seven road maintenance scenarios, assessing alert accuracy, relevance, information overload, prioritization effectiveness, and consistency. Results demonstrate that the system achieves an overall accuracy of 100%, role-content match rate of 82% and information overload rate of 25%, indicating that alerts are effectively tailored to ensure that each user receives comprehensive information necessary to perform their role, while minimizing information overload. Additionally, the system demonstrates a high level of consistency and prioritization accuracy, with an overall low-priority distraction rate of 7% and high-priority miss rate of 4%, reducing unnecessary distractions and ensuring that high-priority alerts are promptly delivered. The findings highlight the potential of LLM-driven digital twin-initiated interactions to optimize information flow, improve response efficiency, and enhance decision-making processes.]]></description>
      <pubDate>Mon, 02 Feb 2026 09:32:47 GMT</pubDate>
      <guid>https://trid.trb.org/View/2632716</guid>
    </item>
    <item>
      <title>Efficient Track Irregularity Prediction Using TimeGAN and BLS-SVR</title>
      <link>https://trid.trb.org/View/2632715</link>
      <description><![CDATA[Monitoring track irregularity is vital for ensuring the safety and comfort of railway transportation. Accurate predictions of track irregularities allow infrastructure managers (IMs) to develop cost-effective maintenance strategies. This paper presents a novel track irregularity prediction method to tackle the challenges of nonlinearity, randomness, and data scarcity in irregularity datasets. The proposed approach enhances data fidelity by leveraging an improved time-series generative adversarial network (TimeGAN) integrated with a stationary gamma process (SGP) and statistical feature constraints to capture temporal dependencies and distributional patterns. Subsequently, Bayesian least-squares support vector regression (BLS-SVR) is employed to predict track irregularity changes over time using a multistep prediction method based on the Monte Carlo algorithm. Experimental results show that the proposed method significantly outperforms minimum-description-length-based rail track deterioration adaptive segmentation framework (MDL-RTDAS), support vector regression (LS-SVR), linear regression (LR), convolutional long short-term memory (ConvLSTM), and temporal convolutional network-long short-term memory (TCNLSTM), achieving average root-mean-square error (RMSE) reductions of 39.03%, 21.46%, 48.24%, 27.36%, and 25.95%, respectively. The corresponding R2 values increased by 42.26%, 14.31%, 47.98%, 9.58%, and 7.60%, respectively. In terms of mean absolute error (MAE), the reductions are 46.76%, 37.77%, 22.04%, 27.24%, and 25.30%, respectively. For mean absolute percentage error (MAPE), the improvement reached 45.22%, 37.94%, 26.50%, 28.42%, and 23.45%, respectively.]]></description>
      <pubDate>Mon, 02 Feb 2026 09:32:47 GMT</pubDate>
      <guid>https://trid.trb.org/View/2632715</guid>
    </item>
    <item>
      <title>Synthetic-to-Real Domain Adaptation with Virtual Laser Scanning and Self-Training–Based Category-Aware Cuboid Mixing for Semantic Segmentation of Bridge Point Clouds</title>
      <link>https://trid.trb.org/View/2630503</link>
      <description><![CDATA[A scarcity of real-world point clouds poses a considerable challenge in training a bridge semantic segmentation model. Although virtual point cloud synthetization offers a promising alternative, the persistent domain gap between synthetic and real-world data remains a critical obstacle. To address this, we present a synthetic-to-real domain adaptation method that integrates virtual laser scanning (VLS) and self-training-based category-aware cuboid mixing (ST-CACM). Our experimental evaluation demonstrates the method’s effectiveness in bridge semantic segmentation through comparison with models trained on real-world point clouds and traditional synthetic point clouds. The proposed approach achieves an overall accuracy of 95.34%, a mean class accuracy of 94.31%, and a mean intersection over union of 89.83%, demonstrating performance comparable to that of the baseline models while significantly reducing the dependency on real-world training data. Notably, both core components, VLS and ST-CACM, effectively mitigated the domain gap between synthetic and real-world data, establishing a robust framework for synthetic-to-real domain adaptation in bridge segmentation tasks. These findings will advance the reconstruction of digital twins and the efficient operations and maintenance of in-service bridges.]]></description>
      <pubDate>Mon, 02 Feb 2026 09:32:47 GMT</pubDate>
      <guid>https://trid.trb.org/View/2630503</guid>
    </item>
    <item>
      <title>Building Fire Evacuation Risk Assessment and Mapping Considering Safety and Travel Delays</title>
      <link>https://trid.trb.org/View/2628248</link>
      <description><![CDATA[Previous studies on building fire evacuation risk often overlook time delays caused by congestion and fire hazards. This gap is addressed by defining travel delays during fire evacuations and introducing two key metrics—delay times and the delay ratio—into the assessment framework. Furthermore, a corresponding parameter calculation method is proposed that considers the characteristics of the occupant trajectory data output from the evacuation simulation. A fire evacuation risk assessment model is developed using a university dormitory as a case study. The model integrates industry foundation classes (IFC) data and advanced technologies, thereby enhancing its applicability to large, multistory buildings. The spatial and numerical distributions of fire evacuation risk indices are subsequently mapped and analyzed in specific scenarios, with risk regions identified on the basis of predefined criteria. In this case, safe regions may incur longer travel delays, whereas unsafe regions may experience fewer delays. Notably, a relatively weak correlation is found between safety and travel delays within the identified risk regions. These findings underscore the importance of considering both safety and travel delay metrics when evaluating fire evacuation risk in buildings, highlighting the need for a more comprehensive approach that accounts for both factors in risk assessments. This research contributes to more accurate building fire evacuation risk assessments and supports the design of effective evacuation strategies and informed emergency decision making.]]></description>
      <pubDate>Mon, 02 Feb 2026 09:32:47 GMT</pubDate>
      <guid>https://trid.trb.org/View/2628248</guid>
    </item>
    <item>
      <title>ConvNextV2X: Road Microcrack Detection Algorithm Based on Improved Multiscale Feature Fusion</title>
      <link>https://trid.trb.org/View/2628076</link>
      <description><![CDATA[Road microcracks are difficult to detect in their early stages. If not detected and repaired in a timely manner, they gradually expand under continuous traffic loads and environmental factors. This expansion eventually can lead to larger cracks or potholes. Such damage severely affects road smoothness and driving comfort. To address this issue, this study proposes an improved multiscale feature fusion algorithm based on the ConvNextV2 network. The algorithm is designed for image processing tasks related to road microcrack detection. Specifically, the algorithm uses four feature calibration layers to adjust the number of input channels. Attention mechanisms then are applied to weight the calibrated features. The weighted features are passed through convolutional layers at different scales. This step extracts multiscale information. Gamma normalization and dropout layers are used to enhance the model’s generalization capability. Global average pooling is employed to compress the spatial dimensions. This process extracts global information from each scale. Finally, the features are concatenated along the channel dimension. This forms a multiscale feature representation as the output. Experimental results show that the proposed algorithm significantly outperforms existing methods in key metrics. Accuracy and precision are improved by 2%–3%. Furthermore, the algorithm achieves precise detection of road microcracks in data sets captured from the perspective of uncrewed ground vehicles (UGVs). This highlights its significant practical application value.]]></description>
      <pubDate>Mon, 02 Feb 2026 09:32:47 GMT</pubDate>
      <guid>https://trid.trb.org/View/2628076</guid>
    </item>
    <item>
      <title>Lightweight Cascade Algorithm for Concrete Crack Detection with Enhanced SSD Techniques</title>
      <link>https://trid.trb.org/View/2628075</link>
      <description><![CDATA[The detection and analysis of structural cracks in concrete bridges remain a critical challenge, particularly for routine inspections of key components such as beams, piers, and pavement. Cracks serve as vital indicators of structural health, yet traditional machine learning-based crack identification methods are often hindered by noise and uneven lighting, resulting in suboptimal accuracy. While general deep learning algorithms for irregular object segmentation, such as cracks, have shown promise, they often produce coarse results and may still require manual intervention, making the process inefficient. To address these limitations, this paper introduces a cascaded machine learning framework that integrates deep learning with morphological algorithms to achieve precise segmentation and extraction of crack regions, followed by accurate computation of crack parameters such as length and width. The proposed approach begins with an enhanced single shot multibox detector (SSD) object detection network to identify and localize cracks within the image. Subsequently, the k-means clustering algorithm and morphological operations are employed to refine the crack morphology. Finally, leveraging the skeleton line and edge information of the crack, an improved orthogonal projection method is utilized for precise parameter measurement. Experimental results demonstrate that the improved SSD algorithm achieves a mAP@0.5 of 91.87 on a publicly available crack dataset, surpassing other state-of-the-art detection algorithms such as YOLO v11 and DETR. Notably, it also excels in computational efficiency, with a training time of 4.1 h and an inference time of just 0.27 s. Furthermore, the proposed method achieves a maximum error of 3.95% (average error of 2.32%) in crack width measurement, meeting the stringent requirements of practical detection scenarios.]]></description>
      <pubDate>Mon, 02 Feb 2026 09:32:47 GMT</pubDate>
      <guid>https://trid.trb.org/View/2628075</guid>
    </item>
    <item>
      <title>Adaptive Analysis for Random Response of Wind–Vehicle–Bridge System Based on a Hybrid Deep-Learning Method</title>
      <link>https://trid.trb.org/View/2625825</link>
      <description><![CDATA[A hybrid method for random response analysis of a wind–vehicle–bridge (WVB) system is presented. The method employs the gated recurrent unit (GRU) network to predict dynamic responses of the WVB system, enhanced by the integration of the sparrow search algorithm (SSA) for autonomously determining the optimal hyperparameters, thus called SSA-GRU. The WVB system is modeled to generate the training data sets, and then the proposed SSA-GRU method establishes a comprehensive mapping between wind excitations and dynamic responses of the WVB system. A long-span railway bridge is adopted as a practical application, and aerodynamic coefficients of the vehicle–bridge system are obtained by wind tunnel tests. Using the data sets, the SSA-GRU method is refined to obtain both the mean value and standard deviation of the system’s dynamic responses. Comparative results with the Monte Carlo method and the traditional GRU method demonstrate a high degree of consistency and significant improvement in computational efficiency.]]></description>
      <pubDate>Mon, 02 Feb 2026 09:32:47 GMT</pubDate>
      <guid>https://trid.trb.org/View/2625825</guid>
    </item>
    <item>
      <title>Tunnel Crack Detection and Evaluation Based on Improved Holistically Nested Edge Detection Network</title>
      <link>https://trid.trb.org/View/2603832</link>
      <description><![CDATA[Crack detection is essential for the routine inspection of subway tunnels. This study addresses key challenges, such as low image acquisition efficiency, poor recognition performance, and the inability to quantify and evaluate cracks. A novel method is proposed for acquiring and detecting subway tunnel crack images, with additional features for parameter calculation and risk assessment. The approach combines deep learning with digital image processing in a tunnel crack detection system, using the multiscale block (MSBlock) module for hollow convolution and an improved holistically nested edge detection (HED) network with a feature pyramid to achieve automatic crack extraction. Postprocessing algorithms, including adaptive thresholding, connected component filtering, and morphological closing, identify crack types and compute parameters like measurements and risk scores, enhancing risk assessment. Based on the acquisition of real metro tunnel images and experimental data analysis, the results show that the proposed algorithm achieves pixel-level detection with a mean IoU of 0.8922 and an accuracy of 92.77%. The algorithm also demonstrates precision rates of 95.06% for crack length, 91.23% for maximum width, and 97.36% for pixel area.]]></description>
      <pubDate>Wed, 05 Nov 2025 10:04:51 GMT</pubDate>
      <guid>https://trid.trb.org/View/2603832</guid>
    </item>
    <item>
      <title>Advancing Traffic Safety Analysis: A Novel Lightweight Rule-Based and Part-of-Speech Tagging-Based Approach for Information Extraction from Crash Reports</title>
      <link>https://trid.trb.org/View/2601400</link>
      <description><![CDATA[Road safety remains a critical issue as traffic accidents continue to rise. Analyzing crash reports is vital for understanding accident causation and implementing preventative measures. In this research, we focused on developing an information extraction system utilizing natural language processing (NLP) to enhance the interpretation of traffic crash reports. While standardized forms offer basic information, the unique details and contexts of each crash require more advanced techniques for comprehensive analysis. We employed a rule-based approach to extract information on unstructured natural language, emphasizing syntactic and light semantic feature recognition in traffic crash narratives. The rule-based approach focused on extracting subjects, actions, and objects of events in crash reports. We prepared a data set of 80 crash reports for training and 20 for testing from Michigan Office of Highway Safety Planning. A new ruleset was developed from the training data, incorporating part-of-speech (POS) tagging and sentence structure patterns matching to extract target information. The extraction process employed the General Architecture for Text Engineering, utilizing its essential NLP resources to find matches of POS tagging features and sentence structures effectively. Experiments on the testing data demonstrated 95.4% precision and 86.9% recall without typos/grammar correction for the data, with improvement to 96.7% precision and 90.16% recall with typos/grammar correction. These results outperformed the state of the art including ChatGPT-4o, highlighting the potential of rule-based NLP techniques by mainly using POS tagging in extracting key information from crash report narratives. This research offers a robust tool for improving road safety analysis.]]></description>
      <pubDate>Wed, 05 Nov 2025 10:04:51 GMT</pubDate>
      <guid>https://trid.trb.org/View/2601400</guid>
    </item>
  </channel>
</rss>