Eyetracking-based guided attention to model cycling perception of safety

Luís Maria Perdigão Bruno

Key information

Authors:

Luís Maria Perdigão Bruno (Luís Perdigão)

Supervisors:

Carlos Jorge Andrade Mariz Santiago (Carlos Jorge Andrade Mariz Santiago); Manuel Ricardo De Almeida Rodrigues Marques (Manuel Marques)

Published in

June 18, 2026

Abstract

Cycling provides substantial benefits for public health, urban mobility, and environmental sustainability, but many individuals avoid it simply because certain urban environments feel unsafe. These subjective safety judgements are heavily shaped by visual cues present in the surrounding urban environment, encompassing both static infrastructure and dynamic traffic elements. Combining street-view imagery with human preferences allows for the scalable modelling of these perceptions to geographically identify which urban layouts are considered safe or unsafe. However, for effective urban analysis, this localization is only the first step; it is essential to understand exactly which visual features drive these assessments and whether computational models rely on the same spatial regions actively inspected by humans. This thesis addresses this interpretability gap by introducing Eye-tracking-Guided Perceived Cycling Safety Network (EG-PCS-Net), a gaze-guided framework for Perceived Cycling Safety. The proposed Siamese Vision Transformer architecture evaluates pairwise street-view scenes to predict safety preferences and assigns a continuous perceived-safety score to each image. To ensure the learned visual evidence accurately reflects human behaviour, eye-tracking data are converted into gaze-derived saliency maps and integrated during training to explicitly supervise the model's internal attention. Experimental evaluations demonstrate that this method maintains competitive predictive performance while generating spatial explanations that are significantly more aligned with human gaze than non-gaze baselines and adapted alternatives. Ultimately, by yielding transparent and perceptually grounded safety assessments, this approach provides urban planners with actionable insights for designing more inviting and secure cycling networks.