SL-YOLO: A Lightweight Road Object Detection Method

The existing road scene object detection algorithms often face problems such as slow inference speed and high computational resource requirements due to high model complexity. This paper proposes a lightweight algorithm SL-YOLO based on YOLOv8 to address these issues. This model uses the StarNet network structure as the backbone, which effectively reduces the computational and parameter complexity of the model. Then, a shared convolution was added to the detection head to construct a lightweight shared convolution detection head, which significantly reduced the computational complexity of the model and improved the performance. In addition, this article also used the LAMP algorithm to prune the final model to reduce redundant computations. The experimental results showed that SL-YOLO outperformed YOLOv8 on both SODA10M and KITTI datasets, with map50 improved by 0.9% and 1.2%, respectively, and the inference speed increased by 36FPS and 41FPS, respectively. In addition, on the SODA10M dataset, the number of parameters and computational complexity were reduced by 2.8M and 5.9GFLOPs, respectively. This indicates that the SL-YOLO algorithm significantly reduces computational complexity while maintaining high accuracy and real-time detection capability in scenarios with limited computational resources.