Comparative Analysis of Traffic Detection Using Deep Learning: A Case Study in Debrecen

This study evaluates deep learning models for vehicle detection in urban environments, focusing on the integration of regional data and standardized evaluation protocols. A central contribution is the creation of DebStreet, a novel dataset that captures images from a specific urban setting under varying weather conditions, providing regionally representative information for model development and evaluation. Using DebStreet, four state-of-the-art architectures were assessed: Faster R-CNN, YOLOv8, DETR, and Side-Aware Boundary Localization (SABL). Notably, SABL and YOLOv8 demonstrated superior precision and robustness across diverse scenarios, while DETR showed significant improvements with extended training and increased data volume. Faster R-CNN also proved competitive when carefully optimized. These findings underscore how the combination of regionally representative datasets with consistent evaluation methodologies enables the development of more effective, adaptable, and context-aware vehicle detection systems, contributing valuable insights for advancing intelligent urban mobility solutions.