Integrating Machine Learning in Pedestrian Forensics: A Comprehensive Tool for Analysing Pedestrian Collisions

Vadhiraj Shrinivas, Christophe Bastien, Huw Davies, Alireza Daneshkhah, Joseph Hardwicke, Clive Edward Neal-Sturgess

Research output: Contribution to journalArticlepeer-review

Abstract

Analysis of pedestrian-to-vehicle collisions can be complex due to the
nature of the interaction and the physics involved. The scarcity of
evidence like video evidence (from CCTV or dashcams), data from the
vehicle's ECU, witness accounts, and physical evidence such as tyre
marks, complicates the analysis of these incidents. In cases with
limited evidence, current forensic methods often rely on prolonged
inquiry processes or computationally intensive simulations. Without
adequate data, accurately estimating pedestrian kinematics and
addressing hit-and-run scenarios becomes challenging. This research
provides an alternative approach to enhancing pedestrian forensic
analysis based on machine learning (ML) algorithms trained on over
3000 multi-body computer simulations with a diverse set of vehicle
profiles and pedestrian anthropometries. Leveraging information such
as vehicle profile, damage, and pedestrian attributes like height and
weight, the ML algorithm estimates essential parameters like vehicle
impact speed, pedestrian gait, crossing speed, and crossing direction.
The proposed ML algorithm was evaluated against real-world data
from the UK Road Accident In Depth Studies (RAIDS) and proved to
be accurate in predicting impact conditions within an error tolerance
of 10%. This ML-based technology provides forensic investigators
with vital pedestrian collision parameters early in the inquiry, enabling
a focused analysis on a reduced collision parameter set. First
responders can swiftly estimate speed characteristics, and forensic
analysts can streamline their investigations, potentially aiding legal
procedures and enhancing post-impact care through the use of this in-situ tool.
Original languageEnglish
Article number24SS-01_0328
Pages (from-to)(In-Press)
JournalSAE Technical Papers
Volume(In-Press)
Publication statusAccepted/In press - 6 Feb 2024
EventSAE World Congress 2024 - Detroit, United States
Duration: 16 Apr 202418 Apr 2024

Fingerprint

Dive into the research topics of 'Integrating Machine Learning in Pedestrian Forensics: A Comprehensive Tool for Analysing Pedestrian Collisions'. Together they form a unique fingerprint.

Cite this