Core Concepts
Current pedestrian detectors exhibit significant bias towards children, highlighting the need for fairness improvements.
Abstract
This study evaluates eight state-of-the-art pedestrian detectors across demographic groups on real-world datasets. Significant fairness issues related to age are uncovered, with children being disproportionately undetected compared to adults. The impact of driving scenarios on fairness is explored, revealing biases towards children and females under low brightness and contrast conditions.
Introduction
Autonomous driving systems' susceptibility to software bugs poses risks to pedestrians and passengers.
Extensive research efforts focus on testing autonomous driving systems for safety.
Preliminaries
Software fairness is crucial in ensuring equal treatment of sensitive attributes in decision-making.
Autonomous driving testing techniques involve simulating real-world conditions for system evaluation.
Experimental Design
Research questions aim to assess overall fairness and fairness under different scenarios.
Statistical analysis using metrics like Miss Rate (MR) and Equal Opportunity Difference (EOD) is employed.
Results
Overall fairness assessment reveals significant bias towards children but balanced detection performance for gender and skin tone.
Different scenarios impact fairness, with lower brightness conditions amplifying bias towards children and females.
Discussion
Fairness-performance trade-off may not hold in pedestrian detection systems, suggesting an optimal balance can be achieved by adjusting image conditions.
Implications
Researchers can explore image editing techniques and multi-objective optimization for fairness improvement.
Developers should prioritize addressing biases in pedestrian detection systems to avoid ethical, reputational, financial, and legal repercussions.
Conclusion
The study highlights significant bias in current pedestrian detectors towards children, emphasizing the importance of addressing age-related bias for fairer autonomous driving systems.
Stats
"Our findings reveal significant fairness issues related to age."
"The undetected proportions for adults are 20.14% lower compared to children."
Quotes
"Fairness issues in autonomous driving systems...can perpetuate discriminatory outcomes."
"It is crucial to prioritize fairness testing in autonomous driving systems."