Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Explainable Object-induced Model through Uncertainty for Automated Vehicles

Shihong Ling, Yue Wan, Xiaowei Jia, Na Du

TL;DR

This work addresses the lack of uncertainty-aware explanations in autonomous driving by extending an object-induced decision model with a Beta-prior-based evidential deep learning framework. It integrates uncertainty through a Dirichlet/Beta formulation, and introduces uncertainty-guided data reweighting and augmentation within a two-phase training scheme. Across the BDD-OIA dataset, the approach yields clear improvements in both action prediction and explanatory rationales, outperforming baselines and demonstrating better reliability in uncertain scenarios. While limitations remain (e.g., lane-disambiguation with Faster R-CNN and generalizability), the two-stage strategy and uncertainty-centric components offer a pragmatic path toward more trustworthy explainable AV systems with potential for multimodal and human-in-the-loop extensions.

Abstract

The rapid evolution of automated vehicles (AVs) has the potential to provide safer, more efficient, and comfortable travel options. However, these systems face challenges regarding reliability in complex driving scenarios. Recent explainable AV architectures neglect crucial information related to inherent uncertainties while providing explanations for actions. To overcome such challenges, our study builds upon the "object-induced" model approach that prioritizes the role of objects in scenes for decision-making and integrates uncertainty assessment into the decision-making process using an evidential deep learning paradigm with a Beta prior. Additionally, we explore several advanced training strategies guided by uncertainty, including uncertainty-guided data reweighting and augmentation. Leveraging the BDD-OIA dataset, our findings underscore that the model, through these enhancements, not only offers a clearer comprehension of AV decisions and their underlying reasoning but also surpasses existing baselines across a broad range of scenarios.

Improving Explainable Object-induced Model through Uncertainty for Automated Vehicles

TL;DR

This work addresses the lack of uncertainty-aware explanations in autonomous driving by extending an object-induced decision model with a Beta-prior-based evidential deep learning framework. It integrates uncertainty through a Dirichlet/Beta formulation, and introduces uncertainty-guided data reweighting and augmentation within a two-phase training scheme. Across the BDD-OIA dataset, the approach yields clear improvements in both action prediction and explanatory rationales, outperforming baselines and demonstrating better reliability in uncertain scenarios. While limitations remain (e.g., lane-disambiguation with Faster R-CNN and generalizability), the two-stage strategy and uncertainty-centric components offer a pragmatic path toward more trustworthy explainable AV systems with potential for multimodal and human-in-the-loop extensions.

Abstract

The rapid evolution of automated vehicles (AVs) has the potential to provide safer, more efficient, and comfortable travel options. However, these systems face challenges regarding reliability in complex driving scenarios. Recent explainable AV architectures neglect crucial information related to inherent uncertainties while providing explanations for actions. To overcome such challenges, our study builds upon the "object-induced" model approach that prioritizes the role of objects in scenes for decision-making and integrates uncertainty assessment into the decision-making process using an evidential deep learning paradigm with a Beta prior. Additionally, we explore several advanced training strategies guided by uncertainty, including uncertainty-guided data reweighting and augmentation. Leveraging the BDD-OIA dataset, our findings underscore that the model, through these enhancements, not only offers a clearer comprehension of AV decisions and their underlying reasoning but also surpasses existing baselines across a broad range of scenarios.
Paper Structure (31 sections, 4 equations, 4 figures, 3 tables)

This paper contains 31 sections, 4 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Explainable Autonomous Driving System
  4. From Modular Pipelines to End-to-End Learning.
  5. Enhancing Reasoning for Autonomous Driving.
  6. Uncertainty Quantification
  7. Dataset
  8. Methodology
  9. Preliminaries
  10. Object Detection Model.
  11. Uncertainty Quantification.
  12. Architecture
  13. Uncertainty Integration
  14. Uncertainty-Guided Training Strategies
  15. Data Reweighting.
  16. ...and 16 more sections

Figures (4)

  • Figure 1: Model Architecture: FC=fully-connected layer, EDL Loss=Evidential Deep Learning Loss Function (Derived using the expected value of the cross-entropy loss over the predicted Beta distribution)
  • Figure 2: Comparison with Original OIA: the color of bounding boxes reflects the rank of importance (red to yellow = most to least important), and the color of actions and explanations reflects the correctness (green=correct, red=incorrect, gray=missing)
  • Figure 3: Comparative Efficacy of Enhanced Models: Bounding box colors indicate importance rank (red to yellow = most to least), while action and explanation colors signify correctness (green=correct, red=incorrect, gray=missing)
  • Figure 4: Sample Failure Cases: the color of bounding boxes reflects the rank of importance (red to yellow = most to least important), and the color of actions and explanations reflects the correctness (green=correct, red=incorrect, gray=missing)