Small Target Vehicle Detection Algorithm Based on Improved YOLOv5s

In recent years, YOLOv5s has achieved notable success in object detection. However, its performance in detecting small objects still requires improvement. This paper introduces three key enhancements aimed at improving YOLOv5s for small object detection. Firstly, a detection head for small objects is added, transitioning from three-scale to four-scale detection, thereby enhancing the model's ability to detect small-sized objects. Secondly, the CSP (Cross Stage Partial) structure is replaced by the C3GC (Cross Stage Partial with Global Context) structure, incorporating global context to boost feature representation and improve small object recognition. Finally, the loss function is optimized to Focal-EIOU (Extended Intersection over Union), better assessing bounding box overlap and improving localization accuracy. Experimental results demonstrate that the improved YOLOv5s model significantly outperforms its predecessor on public datasets, with marked improvements in precision and recall, especially in small object detection. These advancements provide an effective solution for small object detection and valuable insights for future research.