smubiainfrastructure/distracted-driver-detection-model

Driver Distraction Detection — Multi-Model System

Classify driver distraction from a single image using an ensemble of Vision Transformer, ResNet50, and EfficientNet models trained on the AUC Distracted Driver Dataset.

Three tabs:

• —Classification — upload a driver image, get distraction class probabilities from ViT, Stacking, or Blending ensemble
• —Grad-CAM — heatmap overlay showing which regions the model attends to
• —LIME — superpixel-level explanation of the prediction

10 distraction classes: safe driving · texting (left/right) · phone call (left/right) · operating radio · drinking · reaching behind · hair/makeup · talking to passenger

Project by Srividya Ravi Sivashankar, Sonia Poh, and Lynus Phua — NUS BIA Data Associate Programme AY25/26