2018 Research Project

Machine Learning Tools for Informing Transportation Technology Design

Principal Investigator
Missy Cummings
Duke University
View Bio


Summary

Rapid advances in transportation technology, such as advanced driver assistance systems (ADAS), V2X connectivity, and progress in developing autonomous vehicles, have the potential to improve safety by addressing the estimated 94% of vehicle crashes that occur due to human error. While the term, “human error,” implies a single source (i.e., the human operator), research in system safety shows it often represents a complex variety of factors and/or a series of events that combine to create a hazardous situation, leading to a crash. However, using conventional statistical methods to analyze crash data to reveal patterns within these combinations is a difficult and labor-intensive task, especially given the size of crash databases. Therefore, it is difficult for designers to develop technological countermeasures that effectively address the interacting factors that lead to human error, and there is a need to develop objective, analytical strategies to help practitioners understand how these factors interact.

Machine learning (ML) techniques are a useful alternative to traditional analysis that could be helpful in informing design decisions from crash data. ML techniques extract meaningful information from unstructured data through mathematical algorithms that find patterns in large datasets. The results allow people to make informed decisions from datasets that would otherwise be too complex to analyze with more conventional statistical tools.

We propose to systematically investigate how ML techniques can be used to design countermeasures that improve system safety, utilizing a framework that explicitly models a broad set of factors. We will use historical crash data to train our model to identify useful relationships that will inform meaningful recommendations to system designers.

Key activities for this work will include:

  • Identify the data elements used to build the model.
  • Models will be developed using an iterative approach.
  • Once a working model (or models) are completed, they will be run against different data sets to ensure consistency.
  • Following validation of the model, the project team will provide an interpretation of the results that can be used by practitioners to inform design decisions. 

The outcomes of this work will include the design recommendations and a comprehensive analysis of crash data that shows relationships among crash factors that have the greatest impact on system safety for vehicle occupants and vulnerable road users. The results and ML algorithms will be available to practitioners through the project report and publication(s).

This work will also be part of a broader research effort at Duke’s Humans and Autonomy Lab (HAL) to improve the explainability of artificial intelligence techniques. Presently, interpreting the results of a ML algorithm often requires extensive subject matter expertise. HAL is currently investigating how the choice of algorithms affects the interpretability of the results to designers or other decision makers who rely upon the outputs of such analysis to make safety-critical decisions regarding future systems. The formation of the design recommendations from the model results will contribute to this effort.

Project Details

Project Type: Research
Project Status: Active
Start Date: 3-1-2018
End Date: 12-31-2019
Contract Year: Year 2
Total Funding from CSCRS: $45,000