Explainable accident and demand prediction analytics

SOTERIA aims to accelerate the attainment of the European Union (EU)’s Vision Zero through a 
holistic framework of innovative models, tools and services that enable data-driven urban safety 
intelligence, facilitate safe travelling of Vulnerable Road Users (VRUs) and foster the safe 
integration of micro-mobility services in complex environments. In doing so, the project explores 
behavioural characteristics of VRUs and engages communities in different locations across Europe 
in social innovation activities for the co-creation of urban safety solutions. 
Deliverable 2.2 focusses on “Explainable accident and demand prediction analytics". Specifically, 
the theory and application of two types of prediction models employed within key components 
of SOTERIA are explained. Accident prediction models allow to predict accident frequencies on 
road segments as well as intersections and are used to provide necessary input for the safe 
routing app and the provision of risk warnings. The models allow to overcome statistical 
fluctuations in frequencies, identify systematic influencing factors and finally, to assess the safety 
of the road network and identify risky areas. Two different modelling approaches were adopted:
Zero-Inflated Negative Binomial (ZINB) models have been created to overcome the “excess” 
records of zero accidents on sections, a major issue for modelling accidents involving VRUs. These 
statistical models allow transparency on the effects of influencing factors but their prediction 
accuracy is constrained in prediction accuracy by the consideration of only uncorrelated 
influencing factors. Therefore, predictive models based on Artificial Intelligence (AI) are created 
which overcome this restriction. Via methods of explainable AI, strong influencing factors can be 
identified and compared to the findings of ZINB models.
Traffic demand prediction models, in contrast, help to describe traffic movements on a 
macroscopic scale, such as cities. The models presented in this deliverable allow to derive travel 
demand across multiple areas of a city and predict the probability of using certain route options 
for different modes of transport. The models overcome the lack of information on road user
movements within the study area. Only for a small proportion of movements information can be 
retrieved from travel surveys, mobile network data or traffic counts at selected locations. The 
proposed approach fusions different data sources, uses machine learning techniques to predict 
travel mode decisions and finally to derive a complete traffic demand model for the city of Madrid. 
The derived traffic volumes per travel mode can be implemented in the accident prediction 
models as influencing factor to describe the traffic exposure of road users more accurately.
Both model types function as base models in SOTERIA’s key components, particularly, the safe 
routing app and the shared mobility operation data.