Research for
Infrastructure.

From field measurements and laboratory experiments to engineering models and AI. Explore the work behind our infrastructure.

Explore all publications ↗
Aerial imagery from the center’s original survey archive
Field imagery · Center archive

Seeing the assets along our roads

2022 · Vision Transformer + YOLOX + FPN
Transformer and YOLOX architecture with highway image examples.
Original project figure · Panboonyuen et al. · CC BY 4.0 · Source ↗

A road-asset detector combines a pretrained Vision Transformer with YOLOX and a feature-pyramid decoder. The study evaluates detection on Thai highway imagery, including small roadside objects such as kilometer stones.

Phisan Santitamnont & coauthors

Information, 13(1), 5

Inside the model

  1. Input

    Thai highway images

  2. Model

    Vision Transformer + YOLOX + FPN

  3. Output

    Road-asset detections

Evaluated against YOLOv5 variants on the Thai highway corpus.

YOLOX · upstream code ↗

Base detector by the YOLOX authors; this is not the Thai-highway study’s implementation.

Code & model availability

Published method. A dedicated code repository and downloadable weights have not been verified.

Publication source ↗

Learning from highway crash records

2025 · VGG-based CNN + Bayesian optimization
Illustration of crash features encoded as grayscale images, a VGG-based CNN, and severity classification.
Method illustration · Drawn for this website; not experimental results · Source ↗

This study transforms discrete crash features into grayscale images using Weight of Evidence, then trains a VGG-based neural network to classify accident severity. It uses Thai highway crash records from 2011–2023.

Kasem Choocharukul & coauthors

Results in Engineering, 27, 106155

Inside the model

  1. Input

    Crash-record features encoded as images

  2. Model

    VGG-based CNN + Bayesian optimization

  3. Output

    Accident-severity classification

Evaluated with F1-score and AUC, with Bayesian hyperparameter optimization.

Code & model availability

Published method. Public code and weights have not been verified. The authors state that they do not have permission to share the data.

Publication source ↗

Mapping buildings with GeoSAM

2024 · Geospatial Segment Anything Model (GeoSAM)
Illustrative aerial building shapes and extracted footprint polygons.
Method illustration · Drawn for this website; not experimental results · Source ↗

Researchers applied GeoSAM to extract building footprints from UAV true orthophotos with a 5 cm ground sampling distance. The study covers science and engineering areas at Chulalongkorn University.

Phisan Santitamnont & coauthors

Engineering Journal of Research and Development, 35(2)

Inside the model

  1. Input

    UAV true orthophotos · 5 cm GSD

  2. Model

    Geospatial Segment Anything Model (GeoSAM)

  3. Output

    Building footprints

Compared extracted footprints with manually delineated outlines using Intersection over Union (IoU).

SamGeo · cited software ↗

The paper cites SamGeo (Wu & Osco, 2023). This is the upstream package, not the study’s code or data.

Code & model availability

Published application of GeoSAM. A study-specific repository and downloadable weights have not been verified.

Publication source ↗

Finding text in complex scenes

2020 · ResNet-50 + FPN
Synthetic street signs and illustrative text-region polygons and offset borders.
Method illustration · Drawn for this website; not experimental results · Source ↗

A convolutional network predicts text, offset and border masks to locate curved and closely spaced text. Polygon offsetting and border augmentation help separate neighboring text instances in natural images.

Thanarat Chalidabhongse & coauthors

Electronics, 9(1), 117

Inside the model

  1. Input

    Natural-scene images

  2. Model

    ResNet-50 + FPN

  3. Output

    Text-region polygons

Evaluated on ICDAR 2015, 2017-MLT, 2019-MLT and Total-Text benchmarks.

Code & model availability

Published method. Public code and downloadable weights have not been verified.

Publication source ↗

These are published studies by researchers in the directory. Publication years and authorship follow the original papers.

Engineering beyond AI

Field measurements, experiments and analytical models across six areas.

Topic illustrations explain the subject; they are not experimental figures. Each title links to the original publication.

Explore all 77 selected publications ↗

Code you can explore

Public repositories linked to studies in the researcher portfolios.

2025 · Thanarat Chalidabhongse

HER2 · Uncertainty-aware classification

Medical-imaging collaboration. Feature extraction, classifier training and analysis notebooks. MIT license.

More applied research & road-safety activities

2023 · Conference paper

Estimation of Highway Maintenance Cost due to Heavy Truck Traffic

Co-authored by Boonchai Sangpetngam · NCCE 28

The study used HDM-4 and data from Highway 344 to model pavement deterioration and maintenance costs over 20 years under four truck categories.

NCCE 28 · 9 July 2023 ↗

2023 · Conference paper

Development of Model for Predicting the Flexible Pavement Strength

Co-authored by Boonchai Sangpetngam · NCCE 28

The proceedings list research on flexible pavement strength, deflection, structural number and prediction models. The source provides a downloadable paper; its web abstract is not supplied.

NCCE 28 · 9 July 2023 ↗

2025 · Research translation & road safety

ThaiRAP road assessment and capacity building

Kasem Choocharukul · ThaiRAP / Faculty of Engineering, Chulalongkorn University

iRAP's January 2025 report identifies Kasem Choocharukul as ThaiRAP lead and describes road assessment and training with the Bangkok Metropolitan Administration. This is a programme activity record, not a journal publication.

iRAP · 22 January 2025 ↗

Sources checked 8 September 2026. The directory is historical; current center projects and membership require separate confirmation.