A Lightweight YOLOv8n Model With Small‐Target Enhancement for Autonomous Driving in Complex Traffic Scenarios

ABSTRACT Small‐target detection in complex traffic scenarios faces challenges of low accuracy and high computational costs, hindering real‐world embedded deployment. This paper proposes an improved YOLOv8n model that balances high performance with hardware efficiency. Our contributions are fourfold: (1) a composite loss function integrating normalized Wasserstein distance (NWD) loss for small‐target robustness and IoU Loss for large‐target stability; (2) a small‐target enhancement module (SEPNet) that combines SPDConv to prevent feature loss and CSP‐OmniKernel (OK) for cross‐scale feature aggregation; (3) a parameter sharing strategy using a fully shared classification head and a learnable Scale layer for efficient multi‐scale localization; (4) multi‐layer knowledge distillation to boost performance without inference cost. Extensive experiments on the KITTI and BDD100K datasets show the model achieves over 4% higher mAP, 9–10% faster inference, 45% fewer parameters, and 50% lower computational cost than the baseline. Crucially, we demonstrate that the model can be further compressed via post‐training quantization to 1.4 GFLOPs, proving its strong suitability for deployment on mainstream embedded platforms like Raspberry Pi and Jetson Nano, thus providing a practical solution for real‐time traffic perception.