Self-Learning Vehicle Detection Dataset for Urban Environments

This dataset was collected as part of a research study aimed at enhancing vehicle detection algorithms through a self-learning approach tailored for urban environments. The primary objective was to minimize dependency on extensive manual labeling and improve adaptability and effectiveness in dynamic urban conditions. The study utilized urban camera infrastructures to gather real-time traffic data, focusing on a diverse range of vehicle types.

The dataset includes images captured from traffic cameras situated at the intersection of Calle de Alcalá and Calle de Velázquez in Madrid, Spain, operated by the Madrid City Council. Data collection spanned from November 30, 2023, to December 6, 2023, covering daytime traffic between 8:30 hours and 18:00 hours. A total of 770 images were captured at approximately 5-minute intervals.

This dataset specifically targets five vehicle types: buses, cars, motorcycles, trucks, and vans, chosen to encompass a wide range of vehicle sizes, shapes, and functionalities commonly encountered in city traffic. A subset of 134 images was manually labeled, into sets for training, validation (fine-tuning phase), and validation (self-training phase). The remaining 653 images were labeled automatically via the self-learning process proposed in the research.