Object Detection for Intelligent Driving Based on Improved YOLOv5

Developing intelligent driving has great potential to enhance traffic safety and reduce economic losses. However, the existing object detection algorithms suffer from large model parameters and high computational complexity, making them difficult to deploy on resource-constrained devices. Lightweight algorithms often sacrifice model accuracy. In this paper, we propose a lightweight YOLOv5 structure (YOLOv5-ODE), which greatly improves the computational efficiency without losing the feature expression ability, by introducing the GGhostRegNet module with low computational burden and the ECA attention mechanism. We innovatively propose a lightweight attention-downs amp ling module, named DeformCAConv, which can be integrated into different stages of the model for better learning of global features and local irregular features. The experi-mental results demonstrate the effectiveness of our proposed lightweight network. Specifically, compared with YOLOv5s, under 54.4% FLOPs reduction, YOLOv5-ODE improves by 1.3% on mAP@0.5:0.95.