A prediction-based framework for developing integrated pedestrian safety systems under high-risk scenarios.

Pedestrians are among the most vulnerable road users and remain a primary focus of intelligent vehicle safety. In high-risk scenarios where avoidance is physically infeasible, vehicles' front-end structural designs (e.g., active hoods) play a crucial role in injury mitigation. However, existing systems typically rely on contact-based sensors, which leave extremely short time for device deployment. To address this challenge, this study proposes a prediction-based integrated pedestrian safety framework. We constructed a vehicle-perspective, human-in-the-loop virtual reality dataset specifically enriched with high-risk boundary cases (near-misses and collisions). Using this data, we developed a short-horizon forecasting model (combining causal CNN and stacked LSTM), which achieved robust sub-meter accuracy in capturing abrupt pedestrian behavior. Compared with conventional contact-based systems, the proposed prediction-based triggering method provides a nominal 200 ms forecast horizon. This gives the active hood actuator more time margin to ensure it fully deploys before the pedestrian's head makes contact. Integrated simulation results demonstrated that this anticipation enabled the active hood to fully deploy before head contact, thereby realizing HIC reductions of up to 62.1% at 45 km/h (speed-dependent range: 5.9%-62.1%) and, importantly, enabling reversible electric actuation that is not feasible under reactive contact-triggered strategies. An evaluation on 110 real-world near-miss cases further confirmed trigger conservatism, with 0 observed false positives under the offline geometric check. Validation on real-world collision cases demonstrated the framework's robustness and transferability. Overall, this methodological framework supports a paradigm shift from post-impact detection to pre-impact prediction, enabling the next generation of proactive pedestrian protection systems.