Weather-Aware Real-Time Collision Avoidance for Autonomous Vehicles: A Simulation-to-Edge Framework for Safety-Oriented Perception and Embedded Deployment

Real-time collision avoidance for autonomous vehicles requires not only accurate object detection but also reliable interpretation of perception outputs and computationally feasible execution under changing environmental conditions. This paper presents a unified simulation-to-edge framework that jointly evaluates multi-class perception, weather-dependent robustness, safety-oriented decision making, and embedded deployment. A synthetic dataset of 5927 images containing 21,021 labeled instances was generated in the CARLA simulator across four urban environments and five weather conditions (i.e., Clear, Rain, Heavy Rain, Fog, and Heavy Fog), covering five safety-relevant classes: vehicle, pedestrian, construction cone, debris, and crosswalk. A YOLO-based detector is integrated with a deterministic collision-avoidance decision engine that combines lane-corridor relevance, class-specific monocular distance estimation, weather-dependent distance thresholds, temporal confirmation, severity hysteresis, and a weather-conditioned reliability fallback to generate four interpretable driving states: CONTINUE, WATCH, SLOW DOWN, and STOP. On the held-out test set, the detector achieved 98.00% precision. Weather-stratified evaluation showed that performance degradation was concentrated primarily in vehicle and pedestrian detection, while static roadside categories remained comparatively robust. When evaluated against reference decisions generated from CARLA ground-truth object states, the decision engine achieved 93.5% agreement and correctly identified 98.7% (157 of 159) of critical STOP scenarios. The complete perception-to-decision pipeline was executed on an NVIDIA Jetson Orin Nano Super, achieving 24.23 FPS and a mean latency of 41.59 ms, exceeding the predefined real-time targets of 15 FPS and 67 ms. Detection and decision accuracy were validated in simulation, while the embedded experiments demonstrate the real-time feasibility of the complete pipeline. Overall, the contribution of this work is the joint system-level evaluation of weather-dependent perception, downstream safety-oriented decision behavior, and real-time embedded execution within a common simulation-to-edge framework.