Data and data collection methodologies for the development of computational models of AV/VRU interaction and their integration into virtual simulation testing of AV : Deliverable 2.3 in the EC ITN project SHAPE-IT

Report, 2023

Several computational models explaining interactions between AVs and the VRUs pedestrians and cyclists have been developed in SHAPE-IT. For instance, there are now models predicting whether a pedestrian or cyclist will cross or yield at an intersection. Further, interaction models were developed and/or verified using different types of data collected in experiments or 'in the wild'. These data were combined and fed to different algorithms that leveraged machine learning to describe road-user behaviour.

This deliverable address both pedestrian and cyclist interactions with AVs, utilising both naturalistic data and data collected in controlled environments. The former comprised site-based and in-vehicle data collections. The latter included data from several virtual environments (e.g., driving simulators, riding simulators, and pedestrian simulation environments).

The main conclusion of this deliverable is that the potential for computational models of AV/VRU interaction to promote AV safety while reducing the cost and time of AV development is high. However, more data is needed before human behaviour (especially in critical scenarios) is captured precisely and comprehensively enough that their integration into virtual simulations delivers explainable, accurate, and reliable results. This deliverable is rather a stepping stone to be used to define intermediate goals for the eventual development of computational models of AV/VRU interaction and their integration into virtual simulations for safety benefit assessment.

Within SHAPE-IT, ESR3, ESR13, and ESR14 developed everyday-driving models that may be used directly in traffic simulations, while the focus of ESR15 has been on methods related to and applications of counterfactual simulations.