Pedestrian zebra crossings in the Czech Republic (2023-2024), ML detections

Pedestrian zebra crossings in the Czech Republic (bounding boxes) Dataset derived from ČÚZK Orthoimagery of the Czech Republic (2023 and 2024 versions), using a custom-trained YOLOv5lu model based on our own annotated training dataset. Each bounding box detection contains a prediction accuracy attribute, and the dataset includes every detection with an accuracy greater than 50%. Detections were also post-processed, where pedestrian zebra crossings far away from roads (FP detections) were removed. Dataset info: epsg: 5514 extent: Czech Republic date of acquisitions of orthoimagery: 2023 (west part of the country) and 2024 (east part of the country) number of features: 67.255 geometry: polygons (bounding boxes) format: geopackage Known limitations: Some bounding boxes are smaller than the actual pedestrian crossing bounding boxes Duplicate detections may occur at the edge of orthoimagery map sheets Some crossings may not be detected, especially if they are obscured by cars or vegetation Accuracy Accuracy was evaluated by pairing with the open dataset of pedestrian crossings in the city of Brno. It is important to mention that this public dataset contains only crossings on public roads, while our dataset, derived from orthoimagery, contains all visible crossings. Part of the omission is also due to street reconstruction, where the finished streets are not present in the orthoimagery, but the crosswalks are recorded in the Brno city database. Thus, the accuracy of the detections is higher than the given numbers. Number of crossings in the city of Brno (reference dataset): 2.583 Number of detections in the city of Brno: 2.636 Omission: 265 Commission: 349 (most of the detections are correct, see above) For more information, please get in touch with the author of the ML model and the creator of this dataset: [e-mail]