A multimodal input-based occupant injury prediction method for vehicle frontal collisions

Accurately predicting occupant collision response curves can reduce reliance on costly physical testing and simulations, playing a key role in advancing vehicle collision safety. Deep learning has gained significant traction and become a cornerstone of modern vehicle safety technologies owing to its fast prediction capabilities. However, existing models frequently overlook the mechanical properties of restraint systems and key spatial parameters, which are challenging to quantify numerically yet critically influence injury outcomes. To compensate for this, we propose a multimodal method that converts these spatial parameters into robust image features, mitigating subjective and measurement errors. Our resulting model (MOIPM) predicts key injury metrics with >0.9 accuracy in merely 3.684 s, a 384.6-fold speed increase over traditional simulations. Crucially, MOIPM outperforms baseline models that lack spatial parameter inputs across all evaluated metrics, demonstrating the value of incorporating image-based spatial representations. This approach establishes a new technical pathway for intelligent vehicle safety.