This dataset is a collection of crosswalk (zebra crossing) images collected on vehicle-mounted cameras, used to study advanced neural network algorithms for crosswalk detection and analysis of car crossing behavior. The dataset contains 3434 images, including real scenes such as daytime, rainy days, occlusion, deformation, truncation, night, damaged, dazzling, etc. The targets includes two categories: crosswalk and guide arrows. The dataset is divided into 3080 training sets and 354 test sets. In addition, this dataset also provides additional 1770 images labeled with or without crosswalk for testing. The correspoding paper is CDNet: a real-time and robust crosstalk detection network on Jetson nano based on YOLOv5, which can be found in doi: 10.1007/s00521-022-07007-9 (Read Paper). The benchmark model is based on the previous SOTA model YOLOv5 and has undergone multiple improvements such as ROI, SE, NST, SSVM, Synthetic Fog, etc. The open source code can be found on Github-CDNet.