@ericyocam6Shipped: AI Traffic Sign Safety — Generative AI & Computer Vision
An open-source generative AI and computer vision proof of concept exploring how deep learning can improve traffic sign safety assessment and support intelligent transportation infrastructure.
The project uses a Conditional Generative Adversarial Network (cGAN) to synthesize realistic variations of Stop, Yield, and Upward Arrow traffic signs based on retroreflectivity and contrast parameters. A separate Convolutional Neural Network (CNN) is trained on the synthetic images to classify traffic signs as SAFE or UNSAFE according to predefined MUTCD-based criteria.
Built with PyTorch and Google Colab, the framework integrates synthetic data generation, rule-based safety labeling, image processing, neural network training, and model evaluation. The CNN achieves 93.04% test accuracy and 94.67% validation accuracy on generated datasets.
The research demonstrates how generative AI can help address limited training data in transportation safety applications and explores potential uses in roadway infrastructure inspection and autonomous vehicle perception. This is an experimental research prototype, not a validated real-world traffic sign inspection system.
Comments2
Checking Stop and Yield signs for faded reflectivity is a chore every city and county road crew has to do, often by driving around at night. Will you try it next on real sign photos from a local public works department?
Thanks, Ryan. That's a logical next step. The current POC uses synthetic traffic sign images to establish the feasibility of the approach, but testing against real-world images would be essential to validate its effectiveness. Partnering with a local public works department to obtain actual sign images and reflectivity measurements could provide a valuable dataset for further evaluation. It would also help determine how well the model generalizes beyond synthetic data. Appreciate the suggestion!
2 comments
Checking Stop and Yield signs for faded reflectivity is a chore every city and county road crew has to do, often by driving around at night. Will you try it next on real sign photos from a local public works department?
Thanks, Ryan. That's a logical next step. The current POC uses synthetic traffic sign images to establish the feasibility of the approach, but testing against real-world images would be essential to validate its effectiveness. Partnering with a local public works department to obtain actual sign images and reflectivity measurements could provide a valuable dataset for further evaluation. It would also help determine how well the model generalizes beyond synthetic data. Appreciate the suggestion!