Explainable Autonomous Vehicle Decisions (XAVD)

Explainability pipeline for autonomous driving decisions using BDD100K, MetaDrive, and Grad-CAM.

XAVD studies why an autonomous driving model makes specific decisions in traffic-heavy scenes.

Project highlights:

  • Trained a ResNet18-based perception model with BDD100K and evaluated behaviors in MetaDrive simulation.
  • Applied Grad-CAM saliency analysis to trace lane markings, pedestrians, and traffic signs that influenced predictions.
  • Compared model confidence and interpretable attention maps to identify failure modes and decision inconsistency.

This project helped me connect model performance with transparent reasoning for safety-critical AI systems.