Fast Lane Detection Method Based on Improved Multi-branch Residual Network

Lane line detection emerges as a pivotal technology within the autonomous driving perception suite, demanding exceptionally rapid processing speeds. Traditional approaches typically treat lane detection as a semantic segmentation issue, posing challenges for swift and efficient detection. This paper considers lane line detection as a classification problem, employing a residual network augmented with multiple branches to extract image features. At the end of the feature network, a spatial and channel-wise attention network is fused to enhance the model's ability to extract features. The extracted image features are classified by the classification network composed of full-connection layer, and the final lane detection results are obtained. Experimental results show that the accuracy and F1-measure of this algorithm on TuSimple dataset and CULane dataset have reached 95.83 and 70.4 respectively. Compared with the original method, it is improved by 2.96% and 2% respectively. Moreover, the model exhibits an impressive inference speed of 155 frames per second (FPS) on an RTX 3070 GPU for images at a resolution of 288×800. These results affirm the model's dual strengths in rapid inference and superior detection precision.