Abstract:

Lane detection is a remarkable practical application of computer vision technology in the field of transportation. However, existing semantic segmentation network models still face certain challenges such as insufficient accuracy and blurred edges in road semantic segmentation tasks. To address these issues, an improved lane segmentation network architecture based on the UNet model is proposed. First, a dual attention module (DAM) is introduced in the skip connections of the UNet model, which prioritizes the importance of lane lines and effectively reduces noise interference. Additionally, dynamic snake convolution (DSConv) is employed to replace traditional convolution methods, enhancing the network’s lane detection ability. To enhance the comprehensiveness and accuracy of lane detection in underexposed or dark backgrounds, an improved adaptive Gamma correction method is introduced in the image preprocessing stage. Furthermore, atrous spatial pyramid pooling (ASPP) technology is introduced at the end of the encoder to enhance network performance. Experimental results show that this model achieves an accuracy of 98.93% on the TuSimple dataset while meeting real-time requirements. Compared to five other semantic segmentation-based lane detection algorithms, the proposed algorithm demonstrates superior recognition performance, thus validating its effectiveness.

Key words: lane detection, semantic segmentation, attention mechanism, dynamic snake convolution, Gamma correction algorithm