| Abstract: | Perceived walkability depends not only on what is physically present in a streetscape, but also on
what pedestrians actually attend to while moving. This project develops an end-to-end, reproducible
pipeline that predicts segment-level walkability ratings (1–5) by combining mobile eye-tracking with
semantic scene understanding and interpretable machine learning. Walk recordings are captured with
wearable eye-tracking glasses and processed in Tobii Pro Lab to export fixations and a synchronized
scene video. Video frames are semantically segmented to obtain per-pixel class labels and grouped
into eight street-relevant super-categories (e.g., roadway, motor traffic, pedestrian infrastructure,
buildings, greenery, people/bicycles, furniture/signage, sky/water). For each (participant, segment),
we compute (i) attention shares (the proportion of fixation time per super-category), (ii) gaze statistics
(number of fixations, mean duration, total fixation time), (iii) pixel composition (the share of pixels
per super-category from sampled frames), and (iv) simple attention–content alignment indicators.
We train Elastic Net and gradient-boosted trees and assess generalization with a leave-oneparticipant-
out protocol. Model behavior is explained with SHAP at global and local levels, and a
lightweight Streamlit demo exposes predictions, feature attributions, and “what-if” adjustments for
design exploration. On our dataset, higher roadway and motor-traffic presence generally depress
predicted walkability, whereas pedestrian infrastructure and cohesive building frontage tend to improve
it, with greenery and visible human activity showing context-dependent effects. The result is a
compact, auditable system that connects human visual attention, objective scene composition, and
perceived walkability in terms that are actionable for urban design and evaluation.***
Ce résumé présente l’objectif, la méthodologie et les principaux résultats de ce mémoire. Nous
combinons le suivi oculaire mobile et la segmentation sémantique des scènes urbaines afin de relier le
contenu visuel (ce qui est présent) à l’attention visuelle des piétons (ce qui est regardé) et d’expliquer
la perception de la marchabilité des segments de rue. Les modèles (Elastic Net, LightGBM) sont
évalués en validation « laissez-un participant de côté » et des explications SHAP rendent les résultats
interprétables pour les urbanistes. |
