The canton of Vaud built tunnels under a road in Gimel so animals can cross safely, with cameras that fire bursts of photos on motion — day and night — to count the crossings.
The problem
Tallying newts, frogs, and toads by hand is slow and easy to get wrong. The goal: automate it from ~2,000 labeled images — a sliver of the million captured, many of them false triggers from blowing leaves or maintenance visits. Data augmentation stretched the usable set further.
Approach
Since one image can hold several animals, I used a Single-Shot Detector on a pre-trained backbone (ResNet / MobileNet), fine-tuned from COCO 2017 via transfer learning to locate and label each one.
Outcome
Measured by mean Average Precision, the models caught some animals but were far from perfect — more labeled data and architectures like YOLOv3 or R-CNN would close the gap. Still, it showed the tedious manual count is automatable. Built with Dylan Mamié and Jérôme Vial for a HES-SO Lausanne course.