In 2023, 452 human-caused wildfires burned 16.5 million hectares across Canada. With climate change intensifying fire seasons, even minor campsite negligence carries catastrophic consequences. No scalable, automated mechanism exists to enforce fire safety compliance at checkout, leaving enforcement entirely dependent on human diligence.

Challenge

Solution!

Building on prior publications in lightweight object detection, we investigated the feasibility of a computer vision-based campsite checkout system as a capstone research project. We fine-tuned a pre-trained YOLOv4-tiny model on a custom dataset of 616 original DSLR photographs and 174 web-scraped images across hazardous and non-hazardous campsite conditions.

We iterated through three dataset versions, experimenting with mosaic augmentation, input resizing, and a novel dual-label ("safe/hazardous") classification scheme to study its effect on scene discernment. Our hardware analysis validated real-world deployment feasibility, achieving 1.8s inference time within standard smartphone RAM constraints.

Tech Stack Python, Colab, Requests, OpenCv, Darknet, TFLite

Deliverables & Results

  1. Novel annotated dataset (616 original DSLR + 174 web-scraped images across 3 versions) & data mining scripts

  2. Fine-tuned YOLOv4-tiny fire & smoke detection model

  3. Formal academic research paper submitted to Queen's University (CISC 499, 2024), with methodology, dataset design, and experimental findings.