Trustworthy AI Perception Module for Autonomous Driving
Researchers propose a Trustworthy AI perception module for autonomous driving that integrates robustness, explainability, and uncertainty estimation. Building on a transformer-based detector, the module derives explanations from attention mechanisms at inference time and validates them via perturbation-based consistency tests. It also includes uncertainty calibration and robustness-enhancing training methods. Experiments demonstrate faithful saliency behavior, improved robustness, and well-calibrated uncertainty. The work addresses the gap between theoretical Trustworthy AI guidelines and concrete implementations for 3D scene understanding.
Key facts
- Deep Neural Networks dominate Autonomous Driving perception but lack transparency.
- Existing Trustworthy AI frameworks are mostly theoretical for 3D scene understanding.
- The proposed module uses a transformer-based detector.
- Explanations are derived from attention mechanisms at inference time.
- Faithfulness is validated using perturbation-based consistency tests.
- Uncertainty estimation and calibration module is integrated.
- Robustness-enhancing training methods are applied.
- Experiments show faithful saliency, improved robustness, and well-calibrated uncertainty.
Entities
Institutions
- arXiv