Scene Matching-Assisted Adaptive Control of Autonomous Vehicles in CARLA Simulator

This paper presents a practical and lightweight system designed to support robust autonomous driving in adverse weather, especially dense fog. Instead of relying on expensive LiDAR or fragile camera-only setups, our approach combines YOLOv5-based object detection with Vehicle-to-Vehicle (V2V) communication. A key feature of the system is a scene matching method that uses ROI Align to compare feature maps between vehicles and verify visual consistency. This enables a two-stage distance estimation process - starting with an estimate based on V2V velocity and timing data, then refining it using the relative sizes of matched objects. We tested the system in the CARLA simulator under varying fog levels, showing stable performance with over 90% accuracy and real-time processing speeds under 0.3 seconds. The system also includes a fail-safe mechanism that ensures safety in extreme conditions. Our results show that combining camera input with V2V communication can provide reliable cooperative driving even when visibility is severely limited.