To enhance the detection effect in curved lane scenarios,a feature extraction network combining a cascaded improved residual block and a snake serpentine convolution module is proposed to optimize feature extraction,along with a mixed-anchor-based position selection and classification method to accelerate accurate localization and classification.The feature extraction network includes multi-scale feature fusion residual blocks and improved residual blocks that integrate spatial channel reorganization convolution,deeply capturing the fine features of lane lines.The snake serpentine convolution module enhances the network's recognition performance in curved areas by dynamically adjusting the convolution kernels.On the Tusimple dataset,this method achieves an accuracy of 95.88％,with a false detection rate and a missed detection rate of 2.84％and 4.25％,respectively,indicating high detection performance.The frames per second(FPS)is 50.7,meeting the real-time detection requirements of autonomous driving cars.In the curved scenario dataset Tusimple_Curve,compared to the UFLD V2 algorithm,this method achieves a 0.23％improvement in accuracy,and the false detection rate and missed detection rate are reduced by 0.22％and 0.36％,respectively,further confirming its effectiveness in curved scenarios.
维普
万方数据