Enhancing semantic segmentation for urban accessibility using high-fidelity synthetic data

Semantic segmentation of urban scenes is essential for the development of smart cities; however, its effectiveness relies heavily on large, pixel-level annotated datasets, which are particularly scarce for mobility aids. This study aims to enhance semantic segmentation for urban accessibility applications by leveraging synthetic data. The proposed methodology integrates high-fidelity synthetic data generation using Unreal Engine 5.1, automated semantic mask processing, and the training of state-of-the-art segmentation models. A dataset of 5,036 images with pixel-perfect labels across 22 classes, including sidewalks, wheelchairs, and walking aids, was created to support this investigation. Two architectures were benchmarked: a baseline U-Net and DeepLabv3+ with ASPP. Pre-training with synthetic data increased global mIoU from 0.0626 to 0.84 (13.4x) and substantially improved precision, recall, and F1-score (by approximately 6.8x, 9.3x, and 10.4x, respectively). For accessibility-critical classes, motorized wheelchairs achieved an IoU of 0.94, and sidewalks attained a recall of 0.98. Overall, all 22 classes surpassed the deployment threshold ( ≥ 0.75 IoU). These findings demonstrate that synthetic data, combined with imbalance-aware training strategies, provides a viable pathway toward robust semantic segmentation solutions for urban accessibility applications.