Rare and safety-critical traffic situations remain challenging for autonomous driving (AD) because they are poorly represented in common training data and may include objects outside standard detector classes. This paper presents a real-time RGB-LiDAR fusion framework for detecting and reacting to rare traffic situations in CARLA. The approach combines YOLOv8-based RGB perception, bird’s-eye-view (BEV) LiDAR clustering, decision-level fusion, an interpretable rule-based safety agent with hysteresis, and an automatic emergency braking (AEB) override. Fused observations are classified as semantic-geometric detections, semantic-only detections, or geometric-only obstacle candidates, where unmatched LiDAR clusters are treated conservatively as candidate-level physical evidence. The framework was evaluated on three CARLA maps and 3CSim-inspired corner-case scenarios, comprising 19253 frames. On a manually annotated subset of 1200 frames, the full pipeline achieved 93.7% precision, 94.7% recall, and a 94.2% F1-score. The CPU implementation processed one frame in 34.7 ms on average, remaining within the 50 ms budget of a 20 Hz simulation tick.

Rare and safety-critical traffic situations remain challenging for autonomous driving (AD) because they are underrepresented in common training data and may include objects outside standard detector classes. This paper presents a real-time RGB-LiDAR fusion framework for detecting and reacting to rare traffic situations in CARLA (Car Learning to Act), a reproducible simulator for AD research. The approach combines YOLOv8n-based RGB perception, bird’s-eye-view (BEV) LiDAR clustering, decision-level fusion, an interpretable rule-based safety agent with hysteresis, Time-to-Collision (TTC)-aware escalation, and an automatic emergency braking (AEB) override above the CARLA autopilot. Fused observations are classified as semantic–geometric detections, semantic-only detections, or geometric-only obstacle candidates, where unmatched LiDAR clusters are treated conservatively as candidate-level physical evidence rather than confirmed rare objects. The framework was evaluated on three CARLA maps and 3CSim-inspired corner-case scenarios comprising 19,253 frames, with additional weather/lighting stress tests and a public nuScenes mini cross-platform check. On a manually annotated subset of 4800 CARLA frames, corresponding to approximately $24.9\%$ of the recorded CARLA log, the full framework achieved $96.2\%$ precision, $97.3\%$ recall, and a $96.7\%$ F1-score for safety-relevant threat detection. The control experiments show that the fusion-based safety agent reduced unnecessary braking to $1.7\%$ compared with $8.6\%$ for the LiDAR-only baseline and achieved event-level success on the annotated critical intervals. The proposed CPU-only implementation maintained real-time performance, with an average processing time of $34.7\mathrm{ms}$.