Machine Learning-based Vehicle Kinematic Response Prediction for Small Overlap Crash

Machine learning system for predicting small overlap crash response using early-stage design data

Project Overview

This project developed an AI-based prediction system for vehicle kinematic response in small overlap frontal crash tests. The system enables engineers to predict crash behavior during early design stages using only vehicle geometry and specification data, eliminating the need for expensive physical prototypes.

My Role

Technical Implementation Lead

  • Implemented machine learning models for crash kinematics prediction
  • Developed vehicle trajectory extraction pipeline using object detection algorithms
  • Conducted correlation analysis between design parameters and crash response
  • Delivered end-to-end prediction system integrated with vehicle development workflow

Background

AI-driven analysis of vehicle kinematics in small overlap crash scenarios

Small Overlap Frontal Crash Test

In small overlap frontal tests, only a portion of the vehicle’s front end collides with a barrier, creating unique crash dynamics that differ significantly from full-width frontal impacts. This test scenario is critical for passenger safety assessment.

The Challenge

Traditional crash test evaluation faced several limitations:

  • Subjective assessment: Vehicle kinematic response classification relied on developer’s subjective judgment rather than quantitative metrics
  • Lack of predictive capability: No method existed to predict crash behavior during early design stages
  • Limited understanding: Correlation between vehicle design parameters and crash kinematics was poorly understood
  • High development cost: Physical crash tests are expensive, limiting design iteration capability

Technical Solution

Machine Learning-Based Prediction Model

Developed a prediction system that forecasts vehicle kinematic response using only early-stage design information:

Input: Vehicle geometry and specification data (available in early design phase)

Output: Quantitative prediction of vehicle trajectory and kinematic behavior during small overlap crash

Key Technical Components

  1. Automated Trajectory Extraction
    • Implemented object detection algorithms to extract vehicle motion from crash test videos
    • Automated conversion of visual data into quantitative trajectory measurements
  2. Design-Response Correlation Analysis
    • Analyzed relationships between vehicle design variables (geometry, structure, weight distribution) and crash kinematics
    • Identified key design parameters that influence crash response
  3. Predictive Modeling
    • Trained machine learning models to predict crash kinematics from design parameters
    • Enabled quantitative response prediction without physical testing

Impact

Early Design Phase Integration

The system enables crash performance evaluation during early design stages when:

  • Physical prototypes don’t yet exist
  • Design changes can be implemented at minimal cost
  • Multiple design alternatives can be quickly evaluated

Quantitative Decision Making

Replaced subjective judgment with objective, data-driven metrics for:

  • Vehicle kinematic response classification
  • Design optimization guidance
  • Safety performance prediction

Development Efficiency

  • Reduced dependency on expensive physical crash tests
  • Accelerated design iteration cycles
  • Improved understanding of design-performance relationships

The methodology developed in this project demonstrates how AI can transform traditional automotive safety engineering by enabling predictive analysis during conceptual design phases.

Project Duration: May 2021 - May 2022 (1 year) Client: Hyundai Motor Company Organization: Smart Design Lab, KAIST

Reference: 학술논문 링크