AI-Based Animal Detection System for Road Safety
Authors
Sunka, Nikhil
Issue Date
2025.17.12
Degree
Master of Business Administration
Publisher
Dublin Business School
Rights holder
Rights
Open Access
Abstract
Animal-related road accidents are still a big issue, especially on highways and backroads where animals often show up without much warning. Solutions like thermal cameras, wildlife crossings, and sensor-triggered signs do help—but they’re usually expensive, tied to a specific spot, and not easy to roll out across large areas.This project takes a different route—using computer vision and deep learning to detect animals in real-time, with the long-term goal of integrating these detections into GPS systems like Google Maps or Apple Maps to alert drivers on the move. To explore this, I trained two powerful object detection models—YOLOv8m and Faster R-CNN—on a carefully selected 30,000-image subset from the Open Images dataset. The process involved weeks of preprocessing, cleaning, model tuning, and testing. YOLOv8m was optimized for real-time use: lightweight, fast, and surprisingly accurate, Faster R-CNN leaned more toward precision. It took its time, but often gave sharper, more accurate detections—especially when it came to identifying specific animal classes. In terms of performance, YOLOv8m delivered a mAP@50 of 75.3% and held a solid F1 score. Meanwhile, Faster R-CNN performed better on class-level accuracy and produced tighter bounding boxes overall. I ran side-by-side visual comparisons, created heatmaps, analyzed confidence levels, and documented failure cases. The result? Each model brought something unique to the table, and combining their strengths could lead to even better performance in future versions. More than just a technical study, this work shows how AI can actually serve people and wildlife at the same time. It opens the door for smarter, more proactive safety tools on the road—where detecting a deer before it crosses could save a life.