Which cycling environment appears safer? Learning cycling safety perceptions from pairwise image comparisons

TL;DR

This work tackles the challenge of measuring perceived cycling safety at scale by using pairwise comparisons of real-world street-view images and training a Siamese CNN (PCS-Net) with a multi-loss framework that explicitly handles ties. The approach learns from both image content and human choices to predict which environments are perceived as safer, with two network variants for classification and ranking, and a joint end-to-end model. Through Berlin-based data and city-wide extrapolation, the authors demonstrate that PCS-Net can accurately map perceived safety across a city and can capture changes in the built environment over time. The method offers a scalable, data-driven tool for urban planners to prioritize interventions and monitor safety perceptions as cities evolve, while acknowledging limitations in ground-truth validation and cross-city transferability.

Abstract

Cycling is critical for cities to transition to more sustainable transport modes. Yet, safety concerns remain a critical deterrent for individuals to cycle. If individuals perceive an environment as unsafe for cycling, it is likely that they will prefer other means of transportation. Yet, capturing and understanding how individuals perceive cycling risk is complex and often slow, with researchers defaulting to traditional surveys and in-loco interviews. In this study, we tackle this problem. We base our approach on using pairwise comparisons of real-world images, repeatedly presenting respondents with pairs of road environments and asking them to select the one they perceive as safer for cycling, if any. Using the collected data, we train a siamese-convolutional neural network using a multi-loss framework that learns from individuals' responses, learns preferences directly from images, and includes ties (often discarded in the literature). Effectively, this model learns to predict human-style perceptions, evaluating which cycling environments are perceived as safer. Our model achieves good results, showcasing this approach has a real-life impact, such as improving interventions' effectiveness. Furthermore, it facilitates the continuous assessment of changing cycling environments, permitting short-term evaluations of measures to enhance perceived cycling safety. Finally, our method can be efficiently deployed in different locations with a growing number of openly available street-view images.

Figures (10)

• Figure 1: Example of the pairwise image comparison survey. Users are shown two images and asked to select the one they consider safer for cycling, if any.
• Figure 2: Architectures of PCS-Net$_C$ (above) and PCS-Net$_R$ (below) and corresponding losses ($L_C$ for PCS-Net$_C$ and $L_R$ and $L_1$ for PCS-Net$_R$), which, when combined, make up PCS-Net. Layer sizes are shown below each layer.
• Figure 3: Framework used for model validation using human labelled data (semi-realistic and our real-world data). Data is split in training sets to train Models A and B, which are evaluated on the development sets until convergence is met. Models A and B are then validated on unseen data (validation sets). The best performing model is then used for extrapolating perceived scores for unlabelled images at a city-wide scale in Section \ref{['sec:application']}.
• Figure 4: Examples of images ranked by their perceived safety scores. Ranked semi-realistic from Model (top) A and real images from Model B (bottom) with increasing (left to right) perceived safety scores are shown.
• Figure 5: Model accuracy with varying number of average comparisons for different paired models.