RoadScene-Complex dataset

1) Collection background: Covering various types of roads in Kunming, Yuxi and Dali Bai Autonomous Prefecture in Yunnan Province from December 2023 to September 2024, including highways (Kunshi Expressway, Dali Expressway), urban trunk roads (Kunming Beijing Road, Yuxi Hongta Avenue), and rural roads (roads around Dali Xizhou Ancient Town), covering 4 types of weather: sunny, rainy, foggy, and night, and 3 time periods of morning peak, evening peak, and flat peak. 2) Acquisition equipment: two types of hardware are used to ensure data diversity: 1) a vehicle dash cam (360 intelligent cloud mirrors S800, resolution 1920×1080, frame rate 30 fps), which is installed on the front windshield of the car; and 2) a handheld high-definition camera (SONY Alpha 7RII., resolution 3840×2160, frame rate 24 fps), with shooting angles covering head-up and overhead views. 3) Data annotation: Using the LabelImg 1.8.6 tool, the annotation rules follow the PASCAL VOC format, and the annotation targets include 6 types of road core elements: car, person, truck, bicycle, bus, and motorcycle. 4) Data statistics: A total of 7676 images (JPG lossless compression format) are marked, and the total number of labeling boxes is 62943. According to the target pixel size, small targets (<32×32) account for 35%, medium targets (32×32~96×96) account for 48%, and large targets (>96×96) account for 17%. The subsets are divided according to the complexity of the scene: the high-occlusion subset (occlusion rate >50%) accounts for 50%, the high-density subset (target spacing < 20 pixels) accounts for 40% (3070 photos), and the conventional subset accounts for 10% (768 photos). 5) Randomly divide the dataset into a training set, a validation set, and a test set at a ratio of 8:1:1 and ensure that the target size distribution and scene type proportion of each subset are consistent with those of the original dataset to avoid data skewness.