基于轻量级YOLO网络与注意力机制的车道线检测方法OA

Aiming at the current issues in lane detection algorithms based on deep learning,such as insufficient real-time performance,limited global feature modeling capability,and the fact that most related studies remain at the simulation validation stage with a lack of deployment verification for practical embedded systems,this paper proposed a lane detection method based on a lightweight YOLO network and an attention mechanism,and completed the deployment and verification from algorithm design to a real-vehicle system.First,a lightweight Faster-Net was introduced as the backbone network,combined with partial convolution(PConv)modules for structural redesign,which significantly reduced computational complexity while maintaining feature extraction capability.Second,a self-attention mechanism was embedded after the SPPF module of the feature fusion network to enhance the model's modeling ability for the global structure of lane lines and long-range spatial dependencies,while suppressing interference from complex backgrounds.The results show that the proposed method achieves a precision of 96.83%on the CULane dataset with a single-frame inference time of 11.0 ms,and an accuracy of 96.33%on the TuSimple dataset with a frame rate of 90.9 FPS,outper-forming current mainstream algorithms.Finally,the optimized model was deployed on the Jetson Orin NX embedded platform,and functional verification was carried out in a real-vehicle sandbox environment.The results demonstrate that the system exhibits good stability and real-time performance in real-world scenarios.This work not only improves the theoretical performance of lane detection algorithms,but also transitions the technology from dataset validation to practical in-vehicle system implementation,providing a complete and feasible technical pathway for the engineering application of autonomous driving perception technology.