<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>Human Detection Capacity of Vehicle Front Sonar Sensors in Light and Small Passenger Cars and Minivan</title>
      <link>https://trid.trb.org/View/2680734</link>
      <description><![CDATA[Sonar sensor systems have been developed to prevent collisions between vehicles and surrounding objects by employing ultrasonic sensors mounted at the front of the vehicle. These systems warn drivers when nearby obstacles are detected. However, relatively few studies have examined the capacity of sonar to detect humans. This study aims to clarify the human detection capacity of front sonar sensors installed in two light passenger cars (LPC-I and LPC-II), one small passenger car (SPC), and one minivan (MNV). The LPC-I, SPC, and MNV were equipped with center and corner sensors, whereas the LPC-II had only corner sensors. Three volunteers-a child, an adult female, and an adult male-participated in the study. Human detectability was assessed using the "maximum detection distance ratio," defined as the ratio of the maximum detection distance for a volunteer to that for a standard pipe. The results showed that both the center and corner sensors consistently detected front- and side-facing human volunteers. For front-facing human volunteers, the maximum detection distance ratios relative to the pipe were 99-101% (child), 93-101% (adult female), and 98-101% (adult male) for the center sonar sensor, and 99-102%, 94-102%, and 96-100% for the corner sensor. For side-facing human volunteers, the corresponding ratios were 97-100%, 92-97%, and 94-99% for the center sensor, and 95-99%, 91-98%, and 93-98% for the corner sensor. These detection ratios were closely aligned with those of the pipe. These findings suggest that front sonar sensors can effectively detect humans prior to vehicle motion initiation, indicating their potential to reduce low-speed vehicle collisions with nearby pedestrians.]]></description>
      <pubDate>Tue, 19 May 2026 15:12:31 GMT</pubDate>
      <guid>https://trid.trb.org/View/2680734</guid>
    </item>
    <item>
      <title>Research on Bridge Surface Damage Classification Based on Improved Vision Transformer Model</title>
      <link>https://trid.trb.org/View/2703810</link>
      <description><![CDATA[To improve the accuracy of surface damage classification for concrete bridges, this paper proposes an improved model—convolutional neural network-vision transformer (CNN-ViT). First, by replacing the original image block operation with a CNN, the model’s feature extraction capability is enhanced, allowing it to retain more critical information from the image. Second, the introduced local aggregation module dynamically focuses attention on the damaged area. By aggregating local features and fusing contextual information, it enhances feature learning and extraction in the damaged region, thereby improving the model’s accuracy and robustness in identifying fine damage in complex backgrounds. Finally, to verify the model’s effectiveness, ablation experiments were conducted, and its performance was compared with that of other neural network models. Experiment results show that the model achieves an accuracy of 98.7% in real-world concrete bridge surface damage identification, which is 10% higher than that of the original model. Compared with other neural network models, the combination of CNN and the local aggregation module effectively suppresses background noise interference and significantly improves the model’s overall performance, with higher detection accuracy and robustness.]]></description>
      <pubDate>Tue, 19 May 2026 09:02:16 GMT</pubDate>
      <guid>https://trid.trb.org/View/2703810</guid>
    </item>
    <item>
      <title>Investigating Nonmotorist Crash Exposure at Highway–Rail Grade Crossings Using Artificial Intelligence-Based Object Detection and Generalized Linear Count Models</title>
      <link>https://trid.trb.org/View/2703807</link>
      <description><![CDATA[A critical aspect of crash prediction models for highway–rail grade crossings (HRGCs) is crash exposure, which is a measure of train and highway traffic. Although data on motor vehicle traffic (e.g., annual average daily traffic) and train traffic at HRGCs are invariably available, nonmotorist traffic data at HRGCs are not readily available. Current Federal Railroad Administration and other HRGC crash models focus on train and motor vehicle traffic, overlooking nonmotorized traffic. Therefore, there is a need to gather nonmotorized traffic data to improve HRGC crash prediction models. To address this gap, nonmotorist traffic video data were recorded in this study at various urban and suburban HRGCs in Nebraska, followed by the application of an artificial intelligence-based You Only Look Once (version 8) algorithm for automated nonmotorist traffic volume detection. Data on HRGC characteristics, including surrounding area population density and land use, were collected to create a comprehensive HRGC safety database for nonmotorists. Three negative binomial models were estimated to analyze pedestrian, bicyclist, and combined nonmotorist exposure in relation to daily volumes, utilizing physical, dynamic, and temporal characteristics of HRGCs. Results indicated that sidewalks, greater visibility, and cloudy weather conditions were associated with increased nonmotorist traffic volume. Conversely, higher vehicular traffic levels, wet road conditions, low population density, and more traffic lanes correlated with lower nonmotorist traffic. This study established an initial framework for nonmotorist traffic monitoring and identified key environmental and technical challenges in automated detection at HRGCs; based on these findings, recommendations for addressing technical limitations were provided for future research.]]></description>
      <pubDate>Tue, 19 May 2026 09:02:16 GMT</pubDate>
      <guid>https://trid.trb.org/View/2703807</guid>
    </item>
    <item>
      <title>A Crack Detection and Quantification Framework for High-Resolution Images Using Mamba and Unmanned Devices</title>
      <link>https://trid.trb.org/View/2646683</link>
      <description><![CDATA[In structural defects inspection, the quantitative detection of slender cracks remains a significant challenge. Existing methods suffer from low segmentation accuracy for complex boundaries and high computational demands for high-resolution (HR) images, making them unsuitable for the current scenarios where unmanned devices are widely deployed. To address the above-mentioned limitations, a crack detection and quantification framework based on multi-scale convolution-enhanced Mamba (MCMamba) and an HR image calibration method is proposed. The MCMamba is designed based on the Mamba architecture and the calibration method using variable step-size moving least squares is proposed to fit the scale field of HR images, enabling precise crack segmentation and quantification. The MCMamba is trained on an established dataset, and the framework is further field-tested using a climbing robot and Unmanned Aerial Vehicle (UAV), achieving accuracy with less than 10% error for cracks thinner than 0.2 mm. This framework improves crack detection accuracy and demonstrates its advantages in quantifying slender cracks on large-scale bridges in engineering practice.]]></description>
      <pubDate>Mon, 18 May 2026 16:36:43 GMT</pubDate>
      <guid>https://trid.trb.org/View/2646683</guid>
    </item>
    <item>
      <title>An Unsupervised Cross-Domain Method for Bridge Damage Detection Based on Multichannel Symmetric Dot Pattern Feature Alignment</title>
      <link>https://trid.trb.org/View/2646681</link>
      <description><![CDATA[A critical issue for data-driven and machine learning-based damage detection of engineering infrastructures is associated with unlabeled datasets and distribution shifts in cross-domains. To overcome this challenge, this study develops an unsupervised cross-domain method for bridge damage detection based on interclass alignment of time-frequency features extracted from multichannel sensor data. The computational framework was developed based on a deep subdomain adaptation network integrating digital and physical information. Initially, a multichannel symmetric dot pattern was utilized to transform the structural acceleration signals into a comprehensive image. Subsequently, a convolutional block attention module-enhanced ResNet34 (CBAM-ResNet34) was constructed to extract discriminative time-frequency features, where a local maximum mean discrepancy principle was introduced to perform class-conditional alignment across subdomains. Compared with traditional global domain alignment methods, the proposed approach focuses on aligning class-conditional distributions within subdomains to improve the generalization performance with unlabeled datasets. The proposed method was validated on both simulated and experimental datasets collected from a laboratory-scaled steel truss bridge. Furthermore, a case study on the Old ADA Bridge in Japan was presented to demonstrate the robustness and practical applicability of the proposed approach, serving as a benchmark against classic unsupervised methods. The results show that the proposed framework has a substantial improvement in source-to-target transfer recognition performance. Discussions were conducted on the application prospects of the proposed framework for more in-service infrastructures in complex conditions.]]></description>
      <pubDate>Mon, 18 May 2026 16:36:43 GMT</pubDate>
      <guid>https://trid.trb.org/View/2646681</guid>
    </item>
    <item>
      <title>Modular Architecture for Traffic Monitoring Systems Using 3d Lidar Sensors</title>
      <link>https://trid.trb.org/View/2646685</link>
      <description><![CDATA[Urban growth has generated an increasing demand for robust and accurate traffic monitoring systems. Traditional technologies such as inductive loops and computer vision have limitations under adverse environmental conditions and in high-density traffic scenarios. This work proposes a modular architecture for traffic monitoring based exclusively on 3D Light Detection and Ranging (LiDAR) sensors, which stand out for their high accuracy and resilience to light and weather variations. The proposed architecture consists of eight independent modular levels. Its main innovations include optimized methods for background subtraction, real-time detection algorithms using machine learning with subsampling techniques, a multi-object tracking system that preserves vehicle identity in the face of occlusions, and a decision tree-based classifier that assigns vehicle type based on geometric characteristics. The solution was validated in real conditions on the M-13 motorway in Madrid (Spain) over more than 100 h of data recordings, using an instrumented gantry with a 3D LiDAR that supervises two-lane roadway, evaluating critical attributes for traffic management such as vehicle detection, classification, and tracking. The proposed design facilitates scalability and compatibility with various sensors, enabling advanced applications in traffic monitoring, cooperative connected vehicle (vehicle-to-infrastructure) contexts, high-level automated driving, and smart highways.]]></description>
      <pubDate>Mon, 18 May 2026 16:36:43 GMT</pubDate>
      <guid>https://trid.trb.org/View/2646685</guid>
    </item>
    <item>
      <title>Three‐dimensional reconstruction of loose defects in semi‐rigid base layers using enhanced deep learning and point cloud from GPR images</title>
      <link>https://trid.trb.org/View/2646702</link>
      <description><![CDATA[Loose defects in semi-rigid base layers can critically compromise the structural performance and long-term serviceability of asphalt pavements. However, accurate identification and quantitative assessment of such defects remain challenging due to the lack of reliable ground-penetrating radar (GPR) data sets and the limitations of existing detection methods. This paper presents an integrated framework that combines data simulation, deep learning–based segmentation, and 3D reconstruction to address these challenges. First, a high-fidelity synthetic data set was generated using a random medium-based forward modeling approach to represent varying looseness depths and configurations. Second, a modified YOLOv8-seg architecture was proposed, featuring a novel multi-scale feature fusion module (DN module) and a Weighted Intersection over Union (WIoU) loss function to enhance segmentation precision under noisy and complex GPR conditions. It achieved a mean average precision (mAP) of 97.25% and a real-time inference speed of 32.05 frames per second (FPS). Third, a 3D point cloud reconstruction approach based on inverse distance weighted (IDW) interpolation was introduced to restore the spatial morphology of the detected defect regions. Delaunay triangulation was then used to estimate the volumetric extent of the defects, achieving an overall estimation accuracy of 78.07%. Finally, the proposed framework was validated on a full-scale pavement model, confirming its effectiveness in defect detection, morphological recovery, and quantitative assessment. The findings provide a reliable computational tool for pavement condition evaluation and maintenance planning.]]></description>
      <pubDate>Mon, 18 May 2026 16:36:43 GMT</pubDate>
      <guid>https://trid.trb.org/View/2646702</guid>
    </item>
    <item>
      <title>A hierarchical framework for three‐dimensional pavement crack detection on point clouds with multi‐scale abnormal region filtering and multimodal interaction fusion</title>
      <link>https://trid.trb.org/View/2646694</link>
      <description><![CDATA[Early crack detection enables timely maintenance actions, which in turn help extend pavement life and reduce maintenance costs. Traditional 2D detection lacks detail, while 3D detection faces accuracy and efficiency challenges. This paper proposes a hierarchical crack detection framework—F²CrackDet-PCD (crack detection based on point cloud data with filtering and fusion). The framework adopts a pre-filtering and fine segmentation strategy (multi-scale anomaly region filtering [MARF]). First, the MARF uses point cloud characteristics to quickly identify potential crack regions. Then, an orthogonal projection converts 3D data into RGB, depth, and normal images, which are combined by MIF-CrackNet (multimodal interaction fusion) to enhance detection accuracy and robustness. Two datasets were developed: RoadScan-2228, capturing realistic road scenes, and CrackNet-1187, emphasizing densely cracked pavement. Experimental results show that the MARF achieves a recall of about 98% on both datasets. F²rackDet-PCD achieves F1-scores of 75.0 on RoadScan-2228 and 78.2 on CrackNet-1187. F²CrackDet-PCD provides a solution of lane-level 3D point cloud crack detection for large-scale road detection.]]></description>
      <pubDate>Mon, 18 May 2026 16:36:42 GMT</pubDate>
      <guid>https://trid.trb.org/View/2646694</guid>
    </item>
    <item>
      <title>A semantic‐enhanced transformer with adaptive fusion for road damage detection</title>
      <link>https://trid.trb.org/View/2646693</link>
      <description><![CDATA[Road damage detection faces significant challenges including extreme scale variations, complex visual interference from road textures, diverse orientational patterns, and irregular boundaries. This paper proposes a semantic-enhanced and adaptive fusion detection transformer to address these domain-specific challenges through two synergistic innovations. The semantic enhancement attention module exploits distinctive frequency-domain characteristics of road damages through learnable spectral processing, where damaged regions exhibit 50.5% higher high-frequency energy, compared to intact surfaces, enabling effective discrimination between structural defects and background interference. The adaptive information fusion module implements a three-stage progressive architecture: loss-less transmission establishes information integrity across extreme scales through amplitude-aware upsampling and attention-driven fusion; omnidirectional pattern capture via multi-directional convolutions addresses diverse damage orientations; dual-path processing optimizes computational efficiency. Comprehensive evaluation across four datasets demonstrates state-of-the-art performance with significant improvements: 83.4% mean average precision at intersection over union threshold 0.5 on UAV-PDD2023 (+3.4% over previous best), 31.2% on CNRDD (+1.3%), 61.9% on RDD2020 (+3.0%), and 90.2% on nighttime NPD (+0.6%), while achieving superior efficiency with 62 giga floating-point operations, 20 million parameters, and 51 frames per second inference speed for real-time processing.]]></description>
      <pubDate>Mon, 18 May 2026 16:36:42 GMT</pubDate>
      <guid>https://trid.trb.org/View/2646693</guid>
    </item>
    <item>
      <title>UAV-based bridge inspection: RAS-YOLO11 for efficient and accurate multi-class surface defect detection</title>
      <link>https://trid.trb.org/View/2651563</link>
      <description><![CDATA[As key transportation infrastructure, concrete bridges are prone to various types of defects during service, necessitating the use of AI models for efficient detection. Existing object detection methods for surface defect detection in concrete bridges still have key limitations. Some methods select schemes through simple parameter tuning without conducting structural innovations on the model to address practical scenarios, resulting in limited accuracy and complexity. Other models, while incorporating modules such as attention mechanisms to improve accuracy, increase model complexity, making edge deployment difficult, and fail to adapt to devices such as unmanned aerial vehicles (UAVs). To address this, this study constructs a multi-class defect dataset and adopts data augmentation. Based on YOLO11, it integrates innovative feature extraction and downsampling modules into the backbone network, replaces the original neck network with a new one, and proposes a novel model. Experiments demonstrate that the model significantly reduces complexity while improving detection accuracy. After pruning and quantifying the model, it is successfully deployed on resource-constrained edge devices, enabling real-time and efficient recognition of various types of bridge defects. This study achieves dual optimization of accuracy and complexity, overcomes the bottleneck that existing high-precision models are difficult to deploy on the edge, and provides technical support for the application of devices such as UAVs in bridge maintenance.]]></description>
      <pubDate>Mon, 18 May 2026 16:36:42 GMT</pubDate>
      <guid>https://trid.trb.org/View/2651563</guid>
    </item>
    <item>
      <title>Structural Safety Evaluation from Computational Modeling of Unknown Bridges Using LiDAR Point Cloud and Nondestructive Testing Data
</title>
      <link>https://trid.trb.org/View/2703878</link>
      <description><![CDATA[This project aims to convert LiDAR point cloud data into a finite element model of an unknown bridge by integrating steel bars identified from nondestructive testing into structural geometries based on LiDAR point cloud and validating the computational model against a reference model created manually using structural drawings. The aim of this study will be achieved by executing four tasks: (1) Data collection from a bridge using drone-based LiDAR flights and nondestructive testing, such as ground penetrating radar for detection and identification of steel reinforcement grids hidden in concrete members. (2) 	Data processing through registration, noise removal, and down-sampling. (3) Automated finite element model generation by integrating hidden features into structural components with outlining geometry of point cloud and discretizing them. (4) Condition assessment by running the computational model with estimated material properties under overloaded trucks and/or earthquake loads.]]></description>
      <pubDate>Mon, 18 May 2026 17:09:05 GMT</pubDate>
      <guid>https://trid.trb.org/View/2703878</guid>
    </item>
    <item>
      <title>Benchmarking Computer Vision-Based Approaches to Derive Engineering-Oriented Condition from Existing UDOT Assets Data</title>
      <link>https://trid.trb.org/View/2685461</link>
      <description><![CDATA[Condition assessment of how transportation infrastructure supports safe and reliable road and highway operation. Departments of Transportation across the country rely heavily on manual inspections, which are time-consuming and costly. This study evaluated whether modern computer vision (CV) methods can support traffic sign condition assessment along Utah highways. High-resolution roadway images collected using a camera-mounted vehicle were curated and annotated for three sign types (regulatory, warning, and guide) and four defect conditions (fading, delamination, missing letters/symbols, and broken signs) based on the Manual on Uniform Traffic Control Devices (MUTCD) standards. This study compared two different CV algorithms of YOLO11 and RT-DETR for traffic-sign detection and defect classification. Overall, the CV models showed promising performance for defect cases where an adequate number of training data existed. For example, for fading, YOLO11 and RT-DETR achieved 75% F1 on the validation. Binary classification of delamination (i.e., delamination versus no delamination) yielded similar performance for both models (68% F1). In contrast, the models showed poor performance to identify missing letters/symbols due to texture overlap with delamination and a limited number of annotated sign images with such defects. The results suggested that data quality and label definition had a greater impact on model performance than the choice of algorithms for the studied models.]]></description>
      <pubDate>Fri, 15 May 2026 17:01:04 GMT</pubDate>
      <guid>https://trid.trb.org/View/2685461</guid>
    </item>
    <item>
      <title>User-Centered Smart Traffic Sign Implementation Development Study</title>
      <link>https://trid.trb.org/View/2703714</link>
      <description><![CDATA[
Flaggers maintain traffic flow through a work zone area despite a shutdown of lanes by providing temporary traffic control. In terms of occupational safety, flaggers have one of the highest risk jobs in the country, with 41 out of every 100,000 workers killed on the job each year. This project developed and tested technology for automatically detecting and documenting the occurrence of near-intrusions into a flagger-controlled work zone. The project developed a low-cost portable device for automatically tracking vehicle trajectories, detecting potential intrusions, and providing audio-visual alerts to warn any errant drivers who might cause a danger to flaggers and workers in the construction zone. A radar sensor on the device is deployed by using a telescoping pole and collects simultaneous measurements from approaching vehicles in multiple lanes.

Tests were conducted in six real-world traffic scenarios, including work zones at one rural location (Cook County), three urban locations (Saint Paul, White Bear Lake, and Eden Prairie), a synthetic urban zone involving pedestrian crossings (Saint Paul) and one suburban/rural location (Mound). Detailed results and analysis are presented in this report. The results indicate that multiple design iterations have improved the device and enabled it to work reliably – Very few false alarms (if any) are triggered and the intrusion detection curves implemented in the system are verified to work well. The vehicles which were alerted using audio-visual warnings in the last work zone test responded appropriately with a majority of them slowing down in response to the alarms.]]></description>
      <pubDate>Fri, 15 May 2026 14:41:38 GMT</pubDate>
      <guid>https://trid.trb.org/View/2703714</guid>
    </item>
    <item>
      <title>On the applicability of time series anomaly detection methods to real-world traffic volume data</title>
      <link>https://trid.trb.org/View/2663716</link>
      <description><![CDATA[Time Series Anomaly Detection (TSAD or TAD) refers to the automatic and data-driven identification of abnormal segments in time series data, a task that has been studied extensively for decades. Despite recent transformative and novel findings revealed by efforts in this field, the literature on traffic anomaly detection has not yet fully reflected on these emerging trends to draw practical conclusions. In this paper, we focus on the applicability of state-of-the-art and well-established TAD methods to road traffic volume data, making contributions in two main ways. First, given the proven and major contribution of evaluation data to TAD outcomes, we argue that existing anomaly-labeled datasets from transportation and traffic systems require substantial enhancements in terms of both data size and label quality. To address this, we propose a new platform to inspect and label large-scale volume data of urban areas based on its unique characteristics and the latest taxonomy of time series anomalies. Second, based on the established framework, we also formulate the TAD problem in traffic volume data and introduce a discord-based, context-embedded, and light-weight traffic anomaly detection method, named Step-isolated Traffic Discords Discovery (Si-TDD), to address this problem. Benefiting from our labeling platform, AnoLT (Anomaly Labeled Traffic) is presented in this paper for the first time as a comprehensive, open-source, and anomaly-labeled spatiotemporal dataset collected from 147 locations across Melbourne, Australia. Comparative results with more than 20 baselines also indicate that Si-TDD considerably outperforms recent TAD solutions when it comes to traffic volume data, achieving a 67% F1 score with the AnoLT dataset. This paper highlights the key role of incorporating context-related information into existing TAD solutions to boost their effectiveness in traffic anomaly detection, a factor that is often overlooked in the current literature.]]></description>
      <pubDate>Thu, 14 May 2026 17:04:32 GMT</pubDate>
      <guid>https://trid.trb.org/View/2663716</guid>
    </item>
    <item>
      <title>Vertical federated learning for transport mode detection using multi-modality data</title>
      <link>https://trid.trb.org/View/2663710</link>
      <description><![CDATA[Transport mode detection (TMD) is vital for intelligent transportation systems and urban computing. However, most existing approaches rely on data from a single modality, such as GPS trajectories or inertial measurement units. This limits their effectiveness in dynamic real-world scenarios. While some studies have improved transport mode detection by combining multiple modalities, heterogeneities in sampling frequency, spatial precision, signal coverage, and occasional missing-modality conditions, together with the requirement for strict temporal alignment, constrain the fused data to the resolution of the weakest modality and lead to substantial information loss. Furthermore, centralizing data from all modalities is often impractical, as the different data types are held by separate parties that may lack labeled data and may also be unwilling to share their raw data due to privacy or commercial concerns. To address these issues, we propose a semi-supervised vertical federated learning (VFL) framework for TMD that uses IMU, GPS, and mobile phone network data. In this framework, each party independently trains an attention-based autoencoder to encode local features. The hidden features are then sent to a central server for classification, and the classification loss is propagated back for local model updating. Given the high-frequency nature of IMU data, we have designed a dynamic sub-segment sampling strategy to adapt to transport mode detection tasks with different temporal resolutions. In addition, the framework adopts distillation and representation alignment to mitigate the impact of missing or weak modalities, and it supports both single-modality and multi-modality inference. We compared the proposed model with several state-of-the-art models, exploring the effectiveness of the different components of the VFL framework. The results demonstrate that our model consistently outperforms multiple baselines, and the proposed VFL framework effectively enhances local inference performance, which can be extended to various model architectures.]]></description>
      <pubDate>Thu, 14 May 2026 17:04:32 GMT</pubDate>
      <guid>https://trid.trb.org/View/2663710</guid>
    </item>
  </channel>
</rss>