Description of Invention
Efficient and accurate vehicle detection lies at the core of Intelligent Transportation Systems (ITS). Despite advancements in deep learning, existing approaches often struggle with scale variation, complex spatial environments, and occlusions typical in dynamic urban settings. This paper presents a Focusing Diffusion Pyramid (FDP) framework explicitly designed to tackle these challenges. The FDP network, central to the architecture, improves feature extraction by progressively refining multi-scale representations, enabling reliable detection of small and occluded vehicles. Enhancing this capability, a Dynamic Head introduces scale-, spatial-, and task-aware attention to adaptively modulate features across varying detection conditions. Moreover, a modified Powerful Intersection over Union loss significantly improves bounding box localization, delivering more precise and compact predictions.