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

Area of Science:

• Automotive Safety Engineering
• Computational Biomechanics
• Artificial Intelligence in Transportation

Background:

• Predicting Head Injury Criterion (HIC) values is crucial for pedestrian safety but is hindered by complex vehicle 3D geometry and imbalanced HIC data.
• Existing methods struggle to accurately model the intricate spatial features of vehicle front-ends and handle the skewed distribution of impact data.

Purpose of the Study:

• To develop an accurate and robust framework (HICnet) for predicting HIC values in pedestrian-vehicle impacts.
• To address the challenges posed by complex 3D vehicle geometry and the long-tailed distribution of HIC data.

Main Methods:

• HICnet utilizes a PointNet++ backbone for 3D point cloud feature extraction, integrating auxiliary physical attributes via multilayer perceptrons.
• A dual-pronged strategy addresses data imbalance: specialized regression networks for hood and non-hood regions and a probability density-based loss function.

Main Results:

• HICnet demonstrated superior predictive performance compared to baseline architectures across 14 vehicle models.
• The framework showed consistent accuracy and robustness across diverse vehicle geometries and impact scenarios.

Conclusions:

• HICnet provides a reliable and accurate tool for automotive safety evaluation, significantly advancing pedestrian impact analysis.
• The proposed methods effectively handle complex geometry and data imbalance, paving the way for improved vehicle safety designs.