I am a physicist, an AI/ML practitioner based in London, Ontario, Canada. I hold a Master of Science in Applied Physics (Energy, Electrical and Electronic Systems) from the University of Yaoundé I, Cameroon, and am currently a SEED Fellow in the Applied AI and Data Analytics for Business certificate at McGill University's School of Continuing Studies. I am a Canadian permanent resident seeking to join the Free Appropriate Sustainable Technology (FAST) Research Group at Western University's Department of Electrical and Computer Engineering as a Master of Engineering Science (MESc) student.

My research interests lie at the intersection of intelligent systems, sustainable energy, computer vision, and open-source hardware areas that align directly with FAST's mission. I bring a strong foundation in power electronics, deep learning, and real-world AI deployment, developed while working full-time as a secondary school educator for over 11 years.

Degree	Institution	Year	Canadian Equivalency (WES)
Applied AI and Data Analytics for Business Professional Development Certificate (SEED Fellow)	McGill University, School of Continuing Studies	In Progress (Expected 2027)	Professional Certificate
Master of Science (Applied Physics) — Energy, Electrical and Electronic Systems	University of Yaoundé I, Cameroon	2020	Master's degree
Bachelor of Science — Physics	University of Yaoundé I, Cameroon	2015	Bachelor's degree (four years)
First Grade Secondary School Teacher Diploma — Physics Education	University of Bamenda (Higher Teacher Training College), Cameroon	2014	Bachelor's degree (three years)

Graduate GPA: 3.19/4.0  |   Master's Thesis Grade: A

Relevant Graduate Coursework: Power Electronics, Non-Conventional Energy Systems and Energy Geopolitics, Signal Processing, Digital Processing of Signals and Images, Applied Optimization and Genetic Algorithms, Tensorial Calculus, Linear Electronics, Automation, Computer-Aided Design of Circuits and Electronic Systems.

Nkamgan, L., Onanena Guelan, R., Mbous Ikong, J., Bossou, O.V., & Essimbi Zobo, B. (2025). "Close-Combat Weapon Detection in Crisis Zones using YOLOv8." Intelligent and Sustainable Solutions — Chronicle of Computing, OkIP International Conference on Artificial Intelligence Frontiers (CAIF), Oklahoma City, USA, April 2, 2025. DOI: 10.55432/978-1-6692-0011-6_7

This work addresses a critical gap in automated weapon detection systems, which focus almost exclusively on conventional firearms. In crisis-affected regions of Cameroon, the primary threats come from close-combat weapons, machetes and sticks used in dense crowd environments where manual surveillance is impractical.

Key contributions:

• Developed a YOLOv8s instance segmentation model via transfer learning in PyTorch for real-time detection and segmentation of sticks and machetes.
• Curated and annotated a novel image dataset using Roboflow, incorporating diverse lighting conditions, backgrounds, crowd densities, and object orientations from real crisis-zone environments.
• Trained two separate models on AWS SageMaker Studio Lab.
• Implemented data augmentation strategies (90° rotations) to mitigate overfitting on a limited dataset.
• Evaluated using Mean Average Precision (mAP).

Relevance to FAST: This project demonstrates my ability to identify a real-world problem, engineer a dataset from scratch, and deploy deep learning models using cloud infrastructure — skills directly transferable to FAST research in computer vision-based quality inspection for photovoltaic cells, vision-guided monitoring of distributed recycling and 3-D printing processes, and open-source appropriate technology (OSAT) for underserved communities.

University of Yaoundé I, Cameroon (2020)  |  Supervisor: Pr. Olivier Vidémé Bossou  |  Grade: A

• Investigated automated detection of dense crowd formations in video sequences to enable timely security alerts, addressing the ongoing crisis in Northwest and Southwest Cameroon.
• Designed and implemented a crowd density detection pipeline in MATLAB 2018b using Harris Corner Detection for feature point extraction combined with DBSCAN clustering to determine whether a dense crowd was present.
• This thesis laid the foundation for my subsequent deep learning work with YOLOv8, demonstrating a progression from classical computer vision to modern deep learning approaches.

Category	Skills
Programming	Python, MATLAB
Deep Learning / ML	PyTorch, TensorFlow/Keras, Scikit-learn, Transfer Learning, CNNs, Instance Segmentation, YOLOv8
Cloud & MLOps	AWS SageMaker (Pipelines, Studio Lab), AWS Lambda, AWS Step Functions, AWS S3, AWS CloudWatch
Data & Annotation	Pandas, NumPy, Roboflow, Seaborn, Jupyter Notebook
Computer Vision	Object Detection, Instance Segmentation, Harris Corner Detection, DBSCAN Clustering, Image Classification
Other	LaTeX, Git
Languages	English (Fluent), French (Fluent)

• Future AWS Agent Engineer — Udacity | AWS AI/ML Scholar Program (in progress, 2026)
• Future AWS AI Engineer — Udacity | AWS AI/ML Scholar Program (2025)
• Build an ML Workflow with Amazon SageMaker (2025)
• Machine Learning Fundamentals — Udacity | AWS AI/ML Scholar Program (2023)
• AI Programming with Python — Udacity | AWS AI/ML Scholar Program (2022)
• Data Analysis Nanodegree (Scholar) — Udacity (2022)
• AWS Machine Learning Foundations (Scholar) — Udacity (2022)
• Applied Data Science Lab — WorldQuant University (2022–2023)

This project presents a YOLOv8-based real-time object detection and instance segmentation system designed to automatically identify close-combat weapons specifically machetes and sticks in surveillance footage from crisis-affected regions of Cameroon. Unlike existing weapon detection systems that focus almost exclusively on conventional firearms, this work addresses the improvised weapons that are actually prevalent in conflict zones across Central Africa. The system was developed using transfer learning with the YOLOv8s segmentation model, trained on a custom-annotated dataset created with Roboflow, and deployed on AWS SageMaker Studio Lab.

This work was published and presented as an oral paper at the 2025 OkIP International Conference on Artificial Intelligence Frontiers (CAIF) in Oklahoma City, USA, on April 2, 2025.

Full citation: Nkamgan Nsonwang, L.S., Onanena Guelan, R., Mbous Ikong, J., Bossou, O.V., & Essimbi Zobo, B. (2025). "Close-Combat Weapon Detection in Crisis Zones using YOLOv8." In: Tiako P.F. (ed) Intelligent and Sustainable Solutions, Chronicle of Computing, OkIP, CAIF25#10. DOI: 10.55432/978-1-6692-0011-6_7

Cameroon has experienced prolonged unrest, particularly in the Northwest and Southwest regions, characterized by frequent attacks and protests. In these crisis environments:

• The primary weapons are not firearms but machetes and sticks improvised close-combat tools.
• Surveillance cameras have been installed across major cities, but they require constant human monitoring, which is time-consuming, resource-intensive, and error-prone.
• Dense crowd environments make manual identification of weapon-carrying individuals extremely difficult.
• Existing automated weapon detection systems are designed for conventional firearms (guns, knives) and fail to address the weapons actually used in these contexts.

Aspect	Existing Systems	This Project
Weapon types	Firearms (guns, rifles, pistols)	Improvised weapons (sticks, machetes)
Detection type	Bounding box detection only	Detection and instance segmentation
Dataset	Generic weapon datasets	Custom dataset from crisis-zone contexts
Regional focus	Western/developed countries	Crisis-affected regions in Central Africa
Application	General security	Real-time surveillance in active conflict zones

The project follows a standard deep learning pipeline: data collection → annotation → model selection → transfer learning → training → evaluation.

A novel dataset was curated specifically for this project, as no existing dataset contained annotated images of sticks and machetes in crisis-zone contexts.

Image sources:

• Online repositories focusing on crisis-affected regions
• Custom photography sessions capturing real-world scenarios

Dataset diversity: Images were collected with deliberate variation in:

• Lighting conditions — day, night, indoor, outdoor
• Backgrounds — urban streets, rural areas, crowd settings
• Object orientations — different angles, grips, and positions
• Crowd densities — isolated individuals to dense groups

Annotation: All images were annotated using Roboflow, a computer vision annotation platform, with polygon masks for instance segmentation.

Data split:

Subset	Percentage
Training	86%
Validation	7%
Testing	7%

Data augmentation: 90° rotations were applied to mitigate overfitting on the limited dataset.

YOLOv8s (the "small" segmentation variant of YOLOv8) was selected after evaluating multiple architectures. YOLOv8 was chosen for its:

• Anchor-free detection head — eliminates the need for predefined anchor boxes
• Joint detection and segmentation — provides both bounding boxes and pixel-level masks
• Feature pyramid networks — enables multi-scale detection of objects at different sizes
• Darknet-53 backbone with Cross-Stage Partial (CSP) connections for efficient feature fusion
• Superior speed-accuracy trade-off — critical for real-time surveillance applications

YOLOv8 Architecture:

Backbone (Darknet-53 + CSP)
  └── Feature extraction at multiple scales
Neck (PANet + Feature Pyramid Network)
  └── Multi-scale feature fusion
  └── Separates classification and detection heads
Head (Anchor-free)
  └── Object detection (bounding boxes + confidence)
  └── Instance segmentation (pixel-level masks)

The trained models demonstrated promising results in detecting and segmenting close-combat weapons:

• Both models successfully learned to detect and segment their respective weapon classes
• The transfer learning approach proved effective for this specialized domain
• The system operates at speeds suitable for real-time surveillance applications
• Instance segmentation provides pixel-level weapon localization, enabling more precise threat assessment than bounding-box detection alone

This project connects to the FAST Research Group's mission in several important ways:

• The project addresses a real security need in an underserved region using accessible AI tools — embodying the OSAT philosophy.
• The methodology uses free and open-source tools: PyTorch (open-source), YOLOv8 by Ultralytics (AGPL-3.0 license), Roboflow (free tier), and AWS SageMaker Studio Lab (free).
• The custom dataset and methodology are documented for reproducibility, enabling other researchers to adapt the approach for their own contexts.

The technical pipeline developed here — custom dataset creation → annotation → transfer learning → real-time inference — is directly transferable to FAST research areas:

• Solar PV quality inspection: Detecting defects (cracks, hotspots, soiling) in photovoltaic cells using the same YOLOv8 detection/segmentation approach.
• DRAM process monitoring: Real-time visual inspection of 3-D printed parts and recycled filament quality using object detection.
• Agricultural monitoring: Detecting crop health indicators or pest damage for resilient food systems.

Experience with AWS SageMaker for model training demonstrates readiness for scalable, cloud-based research workflows — relevant to FAST's distributed and collaborative research model.

Tool	Purpose	License
PyTorch	Deep learning framework	BSD-3-Clause (Open Source)
Ultralytics YOLOv8	Object detection and segmentation model	AGPL-3.0 (Open Source)
Roboflow	Dataset annotation and management	Free tier available
AWS SageMaker Studio Lab	Cloud-based model training	Free tier
Python 3.x	Programming language	PSF License (Open Source)
MATLAB 2018b	Prior thesis work (crowd analysis)	Proprietary

Built ML models with full EDA, feature engineering, and evaluation pipeline using Python, Pandas, Scikit-learn, and AWS SageMaker. (2023)

Developed and optimized neural networks for digit recognition using PyTorch, Torchvision, and Scikit-learn. (2022)

Implemented CNN-based image recognition using transfer learning (ResNet, VGG, MobileNet) with TensorFlow/Keras; prepared and annotated datasets. (2022)

Laurent Stanix Nkamgan Nsonwang — MSc Applied Physics (Energy, Electrical and Electronic Systems), University of Yaoundé I | SEED Fellow, McGill University | Canadian Permanent Resident

I am drawn to the FAST Research Group because its mission sits precisely at the intersection of my technical background and my values:

• Solar Photovoltaic Technology: My graduate coursework in Power Electronics and Non-Conventional Energy Systems, combined with my AI/computer vision skills, positions me to contribute to research in intelligent fault detection for PV systems, computer vision-based quality inspection for photovoltaic cells, and AI-optimized microgrid control.

• Distributed Recycling and Additive Manufacturing (DRAM): My experience building real-time object detection systems with YOLOv8 is directly applicable to vision-guided monitoring of distributed recycling and 3-D printing processes.

• Open-Source Appropriate Technology (OSAT): Having developed AI solutions for crisis-affected regions in Cameroon with limited resources, I understand firsthand the importance of free and open-source approaches to technology that serve underserved communities. I am committed to open-source research and publication.

• Resilient Foods: My background in data-driven analysis and machine learning can contribute to food security research through predictive modeling and optimization.

As a Canadian permanent resident based near Western University, I am ready to contribute full-time to FAST's research mission.

• Production Associate — Magna International, Ontario, Canada (February 2026 – Present)
• Mathematics and Physics Teacher — Ministry of Secondary Education, Yaoundé, Cameroon (September 2014 – February 2026; 11+ years)
  • Designed and delivered curriculum in physics (mechanics, electromagnetism, electrostatics, waves, optics) and mathematics (calculus, statistics, algebra, trigonometry) at the secondary level.
  • Developed data-driven assessment tools to track student performance across cohorts.
  • Created digital instructional resources integrating simulation and visualization tools.
• Stock Accountant — Agrocam Cameroon S.A Poultry Farm, Foumbot, Cameroon (January 2008 – September 2011)

• Email: nkamgans@gmail.com
• Phone: (226) 559-5446
• Location: 31 Feathers Crossing, St. Thomas, ON, Canada