Indian Toll-Plaza Dataset

Indian Toll-Plaza (ITP) Dataset Description The Indian Toll-Plaza (ITP) dataset is a large-scale, real-world visual dataset designed to support research on toll-plaza detection, tracking, and perception tasks for autonomous driving systems. The dataset addresses the lack of publicly available, task-specific data for toll-plaza environments. Dataset Overview The ITP dataset contains a total of 12,500 high-resolution image frames, comprising: 2,500 manually annotated toll-plaza images, and 10,000 non-toll-plaza images representing diverse road scenes. Scene Diversity and Data Collection To capture the complexity of real-world driving environments, data are collected across: Road types: highways, urban roads, rural roads, and campus roads Geographical regions: North, South, East, West, and Central India Cities and highways: Jaipur, Delhi, Patna, Bhopal, Jalandhar, and surrounding national and state highways Time conditions: morning, evening, night, dawn, and dusk Weather conditions: sunny, cloudy, rainy, foggy, smoky, and snowy Traffic densities: crowded, normal, and low All images are stored in PNG format to preserve visual quality and annotation are saved in TXT file format. Annotation Protocol Manual annotations are performed only for toll-plaza images, focusing on bounding-box level labels suitable for object detection tasks. The remaining non-toll-plaza images are provided as background data to support binary classification, imbalance analysis, and robust detector training. Additionally, a large portion of unlabeled data is included to facilitate semi-supervised and self-supervised learning research. Dataset Splits For effective model training and evaluation, the dataset is divided into training, validation, and testing subsets using a 70% / 20% / 10% split, respectively. The split is performed at the road level, ensuring that all images from a specific road appear in only one subset. This strategy prevents data leakage and enhances model generalization. Two experimental settings are supported: Balanced setup: 2,500 toll-plaza images 2,500 non-toll-plaza images Unbalanced setup: 2,500 toll-plaza images 10,000 non-toll-plaza images These configurations allow systematic evaluation of model robustness under realistic class-imbalance scenarios. Each split maintains diversity across: Road categories (state and national highways) Geographic regions Weather and illumination conditions Benchmark Evaluation The dataset is benchmarked using a YOLOv7-based detection model under both balanced and unbalanced class distributions. Experimental results demonstrate improved precision and reliable toll-plaza detection performance, highlighting the dataset’s suitability for autonomous driving perception tasks. The ITP dataset plays a crucial role in understanding lane-choice behavior, safe stopping, and traffic management at toll plazas. Future Extensions Future versions of the ITP dataset will include: Detailed annotations for traffic signs, lane markings, and road infrastructure Pixel-level semantic segmentation labels Extended annotations to support multi-task learning and scene understanding Additionally, specialized deep neural network-based detectors will be developed to leverage these annotations and further enhance autonomous-driving perception capabilities.