From Visual Perception to Biomechanical Reasoning: Learning Human Body Models for Vehicle Safety

Research Project, 2027 –

This project aims to develop learning-based human body representations that enable future autonomous vehicle systems to reason about occupant-specific injury risk in a crash. While modern vehicles are equipped with rich sensor suites and increasingly autonomous decision-making capabilities, current vehicle safety systems are designed based on a few pre-defined human body sizes and postures neglecting the large inter-individual variability in human anatomy, body composition, and posture. This limitation restricts both the accuracy of injury prediction and the potential for adaptive safety strategies.

The project aims to establish a perception-to-representation-to-generation pipeline that links medical imaging, computer vision, and generative modeling to occupant-specific Human Body Models (HBMs) for autonomous vehicle safety. The project has three main objectives: (1) to learn a coherent, full-body statistical shape model (SSM) from heterogeneous medical imaging datasets with varying annotation types and levels of supervision, (2) to investigate how learning methods can infer occupant-specific morphometric variables from natural 2D images captured by onboard vehicle sensors, and (3) to explore generative modeling as an alternative to traditional Statistical Shape Models (SSMs), enabling conditional generation of anatomically plausible HBMs directly from learned latent variables.

By shifting from hand-crafted modeling pipelines to learned, adaptive anatomical representations, the project contributes new methodological foundations for predictive, occupant-aware autonomous systems. The results are expected to enable more accurate injury prediction models and lay the groundwork for future adaptive vehicle safety systems.