Cycling provides substantial benefits for public health, urban mobility, and environmental sustainability, but many people avoid cycling when certain urban environments feel unsafe. These perceived-safety judgments are strongly shaped by visual cues in the surrounding street scene, including both static infrastructure and dynamic traffic elements.
Street-view imagery and human preference data make it possible to model these perceptions at scale, helping identify which urban layouts are perceived as safer or less safe for cycling. However, prediction alone is not enough for meaningful urban analysis. It is also important to understand which visual regions drive the model's safety assessment and whether those regions correspond to the areas actually inspected by humans.
This repository introduces Eye-Tracking-Guided Perceived Cycling Safety Network (EG-PCS-Net), a gaze-guided framework for perceived cycling safety prediction. EG-PCS-Net uses a Siamese Vision Transformer architecture to compare pairs of street-view cycling scenes, predict perceived-safety preferences, and produce continuous perceived-safety scores.
The key idea is to make the model not only predict like humans, but also look more like humans. Eye-tracking data are converted into gaze-derived saliency maps and used to guide the model's internal attention during training. This improves the interpretability of the learned visual evidence while preserving competitive predictive performance.
By producing transparent and perceptually grounded safety assessments, EG-PCS-Net aims to support urban analysis and help planners better understand which visual elements may make cycling environments feel more or less inviting.
This repository brings together three connected parts of the EG-PCS project:
| Component | Purpose |
|---|---|
| EG-PCS-Net | A gaze-aligned Siamese Vision Transformer framework for perceived cycling safety prediction. |
| EG-PCS Dataset | A released dataset with pairwise street-view comparisons, perceived-safety labels, gaze maps, and sanitized eye-tracking source-session files. |
| Deployment application | A standalone interface for producing perceived-safety scores and visual attention heatmaps from street-level images. |
Together, these components support both model development and interpretable urban perception analysis.
EG-PCS-Net compares two street-view images using shared visual encoders and task-specific prediction branches.
- Ranking branch: predicts which image in a pair is perceived as safer.
- Classification branch: supports discrete safety-preference prediction.
- Attention branch: uses eye-tracking-derived saliency maps to guide or evaluate where the model attends.
The attention branch is the main interpretability component of the framework. It encourages the model's visual evidence to become more aligned with human fixation behaviour, making the safety predictions easier to inspect and compare with human attention.
This project introduces the EG-PCS Dataset, a research dataset for perceived cycling safety from street-level imagery. It contains pairwise street-view comparisons, perceived-safety labels, released gaze maps, and sanitized eye-tracking source-session files.
The dataset includes:
- 13,623 pairwise comparison rows;
- 9,790 released street-level images;
- 2,720 fixation-derived gaze-map files;
- 251 survey participants, including 26 laboratory eye-tracking participants;
- sanitized eye-tracking source sessions for transparency, reproducibility, and further gaze-based experiments.
The dataset is available through Zenodo:
https://doi.org/10.5281/zenodo.21242459
Dataset documentation is provided in docs/dataset, including the dataset README, dataset card, data dictionary, eye-tracking source-session guide, license notice, and citation metadata.
We provide a standalone web interface for running EG-PCS-Net inference, allowing users to obtain perceived-safety scores and visual attention heatmaps from street-level images.
The application is intended to make model inference and visualization easier to inspect without manually running training or evaluation scripts.
This repository is organized into three main parts: the model code, the dataset documentation, and the deployment application.
| I want to... | Go to |
|---|---|
| Understand the project at a high level | This README.md |
| Train or evaluate EG-PCS-Net models | src/egpcs/README.md |
| Change experiment settings | configs/ |
| Understand or reuse the dataset | docs/dataset/README.md |
| Check dataset columns and file paths | docs/dataset/DATA_DICTIONARY.md |
| Read dataset limitations, ethics, and intended uses | docs/dataset/dataset_card.md |
| Understand the eye-tracking source sessions | docs/dataset/EYE_TRACKING_SOURCES.md |
| Run the inference web application | deployment_app/ |
| View project figures and screenshots | docs/ |
The most important folders are:
src/egpcs/— core EG-PCS-Net package with model architectures, data loading, training, evaluation, and command-line tools.configs/— experiment and hyperparameter configuration files.docs/dataset/— documentation for the released EG-PCS Dataset.deployment_app/— web application for model inference and attention-map visualization.docs/— architecture diagrams, screenshots, and other documentation assets.
New researchers should start with docs/dataset/README.md to understand the released data, then use src/egpcs/README.md to train or evaluate models.
Please cite the resources you actually use. If your work uses more than one resource, cite each applicable entry.
Cite the software when you use or adapt the training, evaluation, model, or deployment code from this repository.
@software{perdigao2026egpcssoftware,
title = {EG-PCS: Eye-Tracking-Guided Perceived Cycling Safety},
author = {Perdig{\~a}o, Lu{\'i}s Maria},
year = {2026},
version = {0.1.0},
url = {https://github.com/DinhoDarroz/Eyetracking-based-guided-attention-to-model-cycling-perception-of-safety},
license = {MIT}
}Cite the dataset when you use its images, pairwise comparisons, perceived-safety labels, gaze maps, eye-tracking source-session files, or other released data. The dataset is published on Zenodo under CC BY 4.0.
@dataset{perdigao2026egpcsdataset,
title = {EG-PCS: Eye-Tracking-Guided Perceived Cycling Safety Dataset},
author = {Perdig{\~a}o, Lu{\'i}s Maria and Costa, Miguel and Santiago, Carlos and Marques, Manuel},
year = {2026},
publisher = {Zenodo},
version = {1.1.0},
doi = {10.5281/zenodo.21242459},
url = {https://doi.org/10.5281/zenodo.21242459}
}Cite the paper when discussing the EG-PCS method, experiments, results, or scientific findings.
@inproceedings{perdigao2026learning,
title = {Learning to See Like Humans: Gaze-Aligned Cycling Safety Prediction},
author = {Perdig{\~a}o, Lu{\'i}s Maria and Costa, Miguel and Santiago, Carlos and Marques, Manuel},
booktitle = {Proceedings of the IEEE International Conference on Intelligent Transportation Systems},
year = {2026}
}