HICnet: A two-stage framework based on region partitioning for predicting HIC values from car point clouds

Predicting Head Injury Criterion (HIC) values is essential for pedestrian safety assessment but remains challenging due to the complex 3D geometry of vehicles and the long-tailed distribution of HIC data. To capture the vehicle's spatial structure, we represent the front-end as a 3D point cloud and introduce HICnet, a framework built upon a PointNet++ backbone for geometric feature encoding, augmented with multilayer perceptrons to integrate auxiliary physical attributes. To address the long-tailed distribution, we devise a dual-pronged strategy grounded in a key observation: HIC values exhibit inherent spatial segregation, with high and low values predominantly clustered in non-hood and hood regions, respectively. Motivated by this spatial heterogeneity, we first explicitly partition the problem, employing dual specialized regression networks for hood and non-hood regions to capture their distinct impact patterns. Second, we design a tailored loss function that incorporates probability density estimation to directly mitigate the data imbalance. Extensive experiments on 14 vehicle models demonstrate that HICnet consistently outperforms baseline architectures, achieving top predictive performance across the majority of test cases while maintaining robustness to diverse vehicle geometries. This offers a reliable and accurate tool for practical automotive safety evaluation.