Back to Portfolio

Research / In Depth

Research & Publications

My academic work explores computer vision, intelligent transportation and machine learning, connecting research ideas with practical software engineering.

Featured / Master’s Research

Enhanced Deep Learning Models for Real-Time Traffic Analysis

Combining what a traffic camera sees with how traffic evolves over time to study more reliable traffic analysis in difficult conditions.

The Problem
Rain, fog and low light can reduce the reliability of traffic-analysis models. Object detection and traffic forecasting often operate as separate systems.
The Approach
Combine YOLOv8 object detection with ST-GCN modeling of traffic patterns across space and time, using weather-adaptive fusion.
The Result
A unified research prototype connecting traffic perception and prediction, with attention to model evaluation and practical inference constraints.
My Contribution
Designed the integration between detection and forecasting, implemented the model pipeline and evaluated its behavior under adverse environmental conditions.

Research stack

  • Python
  • PyTorch
  • OpenCV
  • YOLOv8
  • ST-GCN

Architecture overview

Simplified view of the Master’s research approach.

Engineering the Research

Perception and prediction

YOLOv8 provides the visual detection component. A spatial-temporal graph convolutional network (ST-GCN) models relationships across locations and time. The research brings these complementary views into a shared traffic-analysis workflow.

Adapting to environmental conditions

Weather-adaptive fusion connects the detection and forecasting components. The research focuses on how these models work together when environmental conditions make visual analysis less dependable.

From models to software

The engineering work connects model inference with a software pipeline, considering API integration, containerized execution and the constraints of serving traffic-analysis outputs to an application.

The thesis manuscript and source code are private. This page presents a summary of the research approach.

Publications & Research Manuscripts

  1. Published2026

    A Multi Stage Ensemble Framework for Fake News Detection: Integrating Traditional Machine Learning, Deep Learning, and Advanced Feature Engineering

    Missang Mi Aba'a Medy Evrard, Qi Wang

    International Journal of Scientific Research and Management · Vol. 14, Issue 03

    Natural Language Processing · Ensemble Learning · Deep Learning

    A hierarchical ensemble combines traditional classifiers, deep learning and engineered language features to classify fake and real news articles.

    The published study uses text cleaning, TF-IDF and lexical, syntactic, semantic and stylistic features. Its four-stage pipeline combines traditional classifiers, advanced classifiers and sequence models through a calibrated Random Forest meta-learner. Evaluation includes stratified splits, cross-validation, ablation studies and statistical comparisons.

    Published study / Reported evaluation

    Reported accuracy
    99.83%
    Reported F1 score
    99.83%
    Articles in the corpus
    44,898

    Reported in the published study for its calibrated Stage 4 meta-ensemble on binary fake/real news classification. The corpus contains 44,898 articles; accuracy and F1 describe the paper’s evaluation, not performance on unseen news sources.

    View Publication (opens in a new tab)DOI: 10.18535/ijsrm/v14i03.ec05 (opens in a new tab)

  2. Research manuscript2025

    Comprehensive Evaluation of Deep Learning Architectures for Traffic Sign Classification: A Multi-Dimensional Benchmarking Study

    Traffic sign classification study

    Computer Vision · Model Benchmarking · Robustness

    A benchmarking study examining how traffic-sign classifiers generalize across datasets, respond to adversarial inputs and perform on different hardware.

    The study compares CNN and transformer architectures through a shared evaluation framework. It considers cross-dataset generalization, adversarial robustness and inference efficiency together. The public repository organizes training and evaluation code, reproducibility configurations and research material.

    View Implementation (opens in a new tab)

  3. Research manuscript2025

    Leak-Free Machine Learning for Real-Time Urban Crash Severity Prediction

    Urban crash severity study

    Machine Learning · Feature Engineering · Intelligent Transportation

    A study of collision-severity prediction using structured traffic records, with a focus on feature availability, data leakage and repeatable model evaluation.

    The research considers which inputs are available before an incident and how feature choices affect evaluation. The related public Python implementation provides data processing, feature engineering, ensemble-model comparisons and analysis reports using NYC collision records.

    View Implementation (opens in a new tab)
Back to PortfolioDiscuss an AI Project