Fairness in Autonomous Driving: Towards Understanding Confounding Factors in Object Detection under Challenging Weather

Weather conditions significantly amplify detection bias: The study found that challenging weather conditions like fog and darkness don't just reduce overall detection performance—they disproportionately harm detection accuracy for certain demographic groups, meaning some pedestrians become harder for autonomous vehicles to "see" depending on weather severity.

Multiple demographic factors create compounding risks: The research identified that gender, skin tone, and body size all independently affect how well pedestrians are detected, with the implication that pedestrians with multiple marginalized characteristics may face compounded safety risks in autonomous vehicle systems.

Scene composition matters for fairness: The distribution of different demographic groups present in a scene influences detection performance, suggesting that fairness in autonomous driving cannot be evaluated in isolation but must account for real-world pedestrian populations and their interactions with detection systems.

Proximity and context are overlooked confounding factors: Beyond just weather and demographics, factors like how close a pedestrian is to the vehicle and other contextual elements significantly affect detection performance, indicating that fairness analysis requires examining multiple interacting variables rather than single factors in isolation.

Current evaluation methods miss bias problems: Despite high accuracy on standard test datasets, the study reveals that algorithmic bias in pedestrian detection remains "unclear" and previously overlooked, highlighting that traditional performance metrics may mask fairness issues critical to public safety in autonomous vehicle deployment.