University Archives

Potholes are a major issue in Detroit, frequently causing severe damage to vehicle tires and alignments. Since manual road inspections are slow and inefficient, our project aims to automate pothole detection using a portable camera-based system. We trained lightweight YOLO-based object detection models on a GPU-equipped Windows PC using a combination of public pothole datasets. To identify a strong starting point for deployment, we compared multiple training configurations and model variants (e.g., different YOLO model sizes and architectures) using the same validation split. Based on these experiments, we selected the YOLO26n as our preferred pretrained model because it provided the best overall detection performance among the tested options.