Porto J., Ziakopoulos A., Avgeros J., López D.F., Yannis G., “Are telematics-based harsh occurrences associated with street-level visual features? A case study of motorway intersections”, Proceedings of the 10th Road Safety and Simulation International Conference, Naples, 23-26 June 2026.

Road intersections are well-known high-risk locations within transportation networks due to their susceptibility to traffic conflicts. This study investigates whether elements visible in street-level imagery are associated with intersection hazard levels, using telematics-based surrogate safety measures: harsh braking and harsh acceleration events. In this exploratory study, telematics data from 322 trips conducted nearby the city of Gorizia, Italy, were analyzed. A total of 98 motorway junctions were selected based on two criteria: (i) at least one trip crossed the intersection, and (ii) at least one connected road was classified as a motorway or primary road. The number of harsh events within a 50-meter buffer around each junction was selected as the target variable. Street-level imagery was sourced from the crowdsourced platform Mapillary, and image features were segmented using a DeepLabV3+ model trained with Cityscapes data. The Spearman test showed weak correlation between number of harsh events and terrain pixels, and no correlation at all when corrected by the Benjamini-Horchberg false detection rate. However, the log-transformed street variables provided a good fit with a Poisson regression model. Street variables were aggregated into five types: construction, flat, poles and signs, road users and nature, whereas the log-ratio was calculated relative to road pixels, present in all images. Flat pixels showed the most statistical significance amongst the environmental attributes, consistent with the correlation results that showed terrain pixels having the smallest p-value. Adding mean speed and number of trips as explanatory variables as opposed to exposure increased the goodness of fit of the model, resulting to a pseudo-R² of 0.57. Due to the limited number of data points, these findings have limited generalization. However, the proposed framework offers a valuable and transferable approach that can be applied to larger datasets in future image-based research.