Blessing Agyei Kyem

I am a first year PhD student at the Sustainable Mobility and Advanced Research in Transportation (SMART) Lab at North Dakota State University, where I work on Multi-modal AI in Computer Vision and its applications in Transportation and Pavement Engineering. My PhD advisor is Armstrong Aboah.

From 2019 to 2022, I worked at Bismuth Technologies as a Data Scientist where I led their Data Team. I contributed to making data-driven decisions by developing predictive models, optimizing data pipelines, and providing actionable insights that significantly improved product performance and customer satisfaction .

I have a BSc in Civil Engineering at Kwame Nkrumah University of Science and Technology, where I worked in Kenneth Adomako-Tutu's lab. After my undergrad, I was a research assistant at the Transport Research and Education Centre KNUST and a teaching assistant for Augustus Ababio-Donkor (CE 367).

Email  /  GitHub  /  Google Scholar  /  ResearchGate  /  Blog  /  X  /  LinkedIn

News

[12/19/2024]         Our paper titled Weather-Adaptive Synthetic Data Generation for Enhanced Power Line Inspection Using StarGAN has been accepted for publication at IEEE Access

[09/25/2024]         Our research paper titled A novel methodological framework for assessing traffic sign retroreflectivity using Lidar data has been accepted for presentation at the TRB 2025 Conference

[09/22/2024]         Our paper PaveCap: The First Multimodal Framework for Comprehensive Pavement Condition Assessment with Dense Captioning and PCI Estimation has been accepted for presentation at the TRB 2025 Annual Conference Meeting

Research

I'm interested in computer vision, machine learning, deep learning and multi-modal models.

Saam-Reflectnet: Sign-Aware Attention-Based Multitasking Framework for Integrated Traffic Sign Detection and Retroreflectivity Estimation

This research involves designing a deep learning framework, called Saam-ReflectNet, that combines traffic sign detection, classification, and retroreflectivity estimation into a single workflow. It utilizes RGB imagery and LiDAR intensity data with an attention-based backbone to enable scalable and accurate traffic sign assessments.

Context-CrackNet: A Context-Aware Framework for Precise Segmentation of Tiny Cracks in Pavement images

This study introduces Context-CrackNet, a novel architecture combining Region-Focused Enhancement Module (RFEM) and Context-Aware Global Module (CAGM) to improve crack segmentation. It outperformed state-of-the-art models across ten datasets in accuracy and efficiency, enabling real-time pavement monitoring.

Weather-Adaptive Synthetic Data Generation for Enhanced Power Line Inspection Using StarGAN

This research introduces a framework combining heuristic image processing and StarGAN to generate synthetic power line inspection images under adverse weather conditions like fog, rain, and nighttime. The study demonstrates that integrating these synthetic images with real data enhances the performance of object detection models, improving robustness and accuracy under diverse scenarios.

PaveCap: The First Multimodal Framework for Comprehensive Pavement Condition Assessment with Dense Captioning and PCI Estimation

We show how to provide dense caption describing the condition of a pavement image. The textual description captures the types of pavement distress present with their severity, the absent pavement distresses and the pavement condition index (PCI)

Advancing Pavement Distress Detection in Developing Countries: A Novel Deep Learning Approach with Locally-Collected Datasets

This research presents a deep learning approach combining YOLOv5 with Convolutional Block Attention Module (CBAM) to detect and classify pavement distress types in developing countries. In addition, we develop a web-based platform to enable real-time pavement distress detection for these developing countries.

Design and source code from Jon Barron's website