Context-Aware Deep Learning based Indian Footpath Damage Segmentation Dataset for Risk Assessment

Damaged footpaths, including cracks and surface irregularities, pose significant risks to pedestrians in urban environments. To facilitate research in automated infrastructure monitoring and pedestrian fall-risk mitigation, we introduce the Indian Footpath Damage Segmentation Dataset, a high-resolution image collection captured using a professional SLR camera under diverse lighting and weather conditions in Pune, India. Each image has been manually annotated with pixel-level segmentation masks using the LabelMe tool, targeting a single class: footpath damage. Annotations include severity levels (low, medium, high) based on crack width, defined in consultation with civil engineering experts. To demonstrate the dataset's utility, we trained a U-Net model with an EfficientNet-B3 encoder and attention-based decoder, and further experimented with ResNeXt50_32x4d encoders and ensemble methods incorporating a convolutional fusion head. Our best-performing ensemble model achieved a Dice score of 0.6899, Intersection over Union (IoU) of 0.6741, accuracy of 0.9317, precision of 0.7154, recall of 0.6932, and F1-score of 0.6899. The dataset is openly accessible for academic and research use on Zenodo.