Research on vehicle detection algorithm based on YOLOv5s-ShuffleNetV2-SE-EIOU

Vehicle detection in drone aerial imagery presents challenges such as complex background environments, small target sizes, and computational complexity, which can result in missed or false detections. To address these issues, an improved YOLOv5s-ShuffleNetV2-SE-EIOU algorithm is proposed. The backbone network is optimized using the lightweight ShuffleNetV2 module, which reduces network redundancy while increasing detection speed. The SE attention module is introduced into the backbone to enhance the model's ability to capture vehicle features, thereby improving detection accuracy in complex environments. The loss function is replaced with EIOU_loss to accelerate the convergence of prediction boxes and improve regression accuracy. Experimental results show that in vehicle detection ablation experiments, this algorithm improved mAP@0.5 detection accuracy by 2.9%, mAP@0.5:0.95 detection accuracy by 3.8%, reduced training time by 0.015 hours, and decreased parameter count by 23.4%. In special weather detection experiments, the algorithm demonstrated higher accuracy.