NeuralMetric: An accurate and efficient real-time safety metric for automated driving systems

Real-time safety metrics are essential for Autonomous Vehicles (AVs) to enhance decision-making, alert human operators, and evaluate safety performance. An effective safety metric must ensure both accuracy and efficiency, reliably identifying imminent safety-critical events while satisfying stringent real-time computational requirements. However, existing methods often rely on heuristic behavioral assumptions and computationally intensive optimization, limiting both their accuracy and real-time applicability. To address these challenges, we propose NeuralMetric, a deep learning-based safety metric that learns to assess situational safety directly from data. NeuralMetric captures complex interactions between the Subject Vehicle (SV) and surrounding Background Vehicles (BVs) without relying on any hand-crafted rules, delivering accurate and adaptive safety assessments. Its Transformer-based architecture ensures exceptional computational efficiency by processing all agents in a single forward pass of the model. Validated across diverse driving environments, including highways, intersections, and roundabouts, NeuralMetric consistently outperforms existing methods in both Precision-Recall (PR) and Receiver Operating Characteristic (ROC) analysis. More importantly, it achieves inference speeds as low as 2 ms, significantly surpassing prior methods and meeting real-time processing demands. These results underscore NeuralMetric’s potential as a reliable and scalable safety metric that advances the safety performance of AVs.