REDESIGNING URBAN STREETS FOR WOMEN’S SAFETY: A PERCEPTION-DRIVEN OPTIMIZATION APPROACH

Pedestrian safety perception shapes walkability, yet existing perception-prediction models rely entirely on street-view imagery, offering no direct link to planning features that can be measured and modified to improve perceived safety. Extending prior work that combines 2D and 3D LiDAR data for fine-scale urban assessment, we test whether measurable LiDAR-derived 3D urban-form features can explain and predict the perceived-safety score produced from 2D image features, and quantify the intervention needed to improve it. The study area is the Azad Market locality of Bhopal, India, using mobile-LiDAR point clouds and corresponding street-level panoramic images. From the point cloud, 310 road segments (25m each) are extracted and matched to 360-degree panoramas divided into views at 0, 120, and 240 degrees. The base perception model, reused from Chauhan et al., was trained on the MIT Place Pulse 2.0 dataset. Eight urban-form features are derived per segment: road width, building height (low/mid/high-rise), vegetation height (low/medium/high-tier), and street-light count. Eight parametric regression models are compared using nested 10-fold x 5-repeat cross-validation. Road width and high-canopy vegetation height show the strongest correlations with perceived safety, while other features show weaker associations. A Generalized Additive Model outperforms the other seven models. Counterfactual optimization estimates the minimum feature change required to improve perceived safety, achieved for the majority of segments. This study demonstrates that perceived-safety scores, when linked with LiDAR-derived urban-form features, can support actionable planning guidance, providing practical decision support for targeted street redesign in Indian cities.