AI × Autonomous Systems
Adverse-Weather AV Perception Benchmark in CARLA
A controlled CARLA benchmark comparing YOLOv8, DETR and Faster R-CNN for pedestrian perception and downstream braking behavior across five weather conditions.
3
Models
5
Weather conditions
YOLOv8n
Practical winner
Closed-loop
System
Question
Which detector produces the most useful pedestrian-perception behavior when visibility degrades — not just the most detections, but the best balance of misses, false positives and safe braking?
Experiment
The same fixed pedestrian-crossing scenario is repeated under Clear Noon, Heavy Rain, Dense Fog, Night and Rainy Night. A remote detector server receives frames and returns detections to a decision layer that can monitor or brake.
Evaluation
The benchmark records detection and miss rate, false-person detections, first braking distance, minimum pedestrian distance and braking behavior. A desktop launcher and results dashboard make experiments reproducible and comparable.
Result
Faster R-CNN often achieved strong recall but produced substantially more false-person detections. YOLOv8n gave the cleanest practical behavior and was selected as the most trustworthy real-time prototype model despite not always having the highest raw detection rate.
Evidence