Mitigating Out-of-Distribution Challenges in Connected Vehicle-Infrastructure-Pedestrian Systems for Promoting Pedestrian Safety at Intersections

Details

• Creators:
  Xu, Susu ; Wang, Chenguang
• Corporate Creators:
  Stony Brook University. Department of Civil Engineering ; Johns Hopkins University ; Rural Safe Efficient Advanced Transportation (R-SEAT) Center
• Corporate Contributors:
  United States. Department of Transportation. University Transportation Centers (UTC) Program ; United States. Department of Transportation. Office of the Assistant Secretary for Research and Technology
• Subject/TRT Terms:
  • Autonomous vehicles
  • Trajectory
  • Artificial intelligence
  • Machine learning
  • Pedestrian safety
  • Traffic mitigation
  • Connected vehicles
  • Infrastructure
  • Intersections
  • Video
  More +
• Resource Type:
  Tech Report
• Geographical Coverage:
  United States
• Edition:
  Final Report: 02/22/2023 - 12/31/2024
• Corporate Publisher:
  Rural Safe Efficient Advanced Transportation (R-SEAT) Center
• Description:
  Pedestrian trajectory prediction constitutes a critical component of motion planning in autonomous vehicles (AVs), facilitating safe and efficient navigation. While recent advancements in deep learning have enabled near real-time prediction capabilities, ethical concerns have emerged within the artificial intelligence (AI) community regarding algorithmic bias and fairness. In particular, the majority of publicly available pedestrian trajectory datasets predominantly represent majority populations, raising concerns about the generalizability of trained models to diverse pedestrian groups, especially vulnerable populations such as individuals with disabilities, older adults, and children. Such biases in prediction models may disproportionately impact these groups, potentially increasing their risk of involvement in vehicular accidents. In this study, we evaluate two state-of-the-art pedestrian trajectory prediction models for potential biases associated with age and gender, utilizing three distinct datasets. We introduce novel evaluation metrics tailored to assess differential model performance across demographic subgroups. Our empirical findings reveal that both models demonstrate reduced predictive accuracy for children and older adults relative to adults, while exhibiting comparable performance across male and female pedestrians. We further analyze potential sources of observed biases and delineate key limitations of our study. Future work will focus on expanding model evaluations, refining bias assessment methodologies, disentangling dataset-related biases from model-induced biases, and developing strategies for mitigating algorithmic bias.
• Format:
  PDF
• Funding:
  69A3552348321
• Download URL:
  https://rosap.ntl.bts.gov/view/dot/94080/dot_94080_DS1.pdf
• File Type:
  [PDF - 1.26 MB]
• Collection(s):
  Autonomous Vehicles ; University Transportation Centers
• Main Document Checksum:
  urn:sha-512:2a1428503372b1922ed61381b5ab9269554a4835b3f56812c2bd6568e9b393c332468afe0024e15e6233a35204159fba53b52be3d972cefbca2b218f8c4ea84e