Research on real-time vehicle recognition for unmanned aerial vehicles based on YOLOv5

To address the limitations of traditional vehicle detection algorithms under UAV perspectives, including high sensitivity to target scale variations, weak robustness against background interference, and insufficient hardware deployment efficiency in complex dynamic scenarios, this study proposes a lightweight vehicle detection method based on the YOLOv5 deep learning framework[1]. By constructing a multi-scenario UAV aerial image dataset and integrating adaptive anchor box optimization strategies with cross-scale feature fusion mechanisms, the proposed approach effectively enhances the model's semantic representation capability for vehicle targets in aerial imagery. Experimental results demonstrate that the bidirectional feature pyramid network (FPN+PAN) enables multi-scale feature interactions, significantly improving small target detection accuracy and adaptability to complex illumination conditions while maintaining stable detection confidence in dense traffic scenarios. The research validates the technical advantages of the YOLOv5s model in balancing detection speed and accuracy on UAV embedded platforms, providing an extensible solution for real-time dynamic target monitoring in intelligent transportation systems, with notable engineering application value.