Abstract
With the rapid development of automatic driving technology, lane image semantic segmentation plays an increasingly important role in intelligent transportation systems. In this paper, a lane image semantic segmentation technology based on the BiSeNetV2 network is proposed. First, we describe the dual-branch structure and feature fusion module in the BiSeNetV2 network, and then elaborate on our improvements in the lane image semantic segmentation task. We incorporated the attention mechanism to help the model grasp the overall structure of the image more effectively and enhance the segmentation accuracy. Simultaneously, we introduce depth separable convolution to decrease computational redundancy and simplify the model's complexity. Ultimately, we performed experiments on the Cityscapes dataset, and the results revealed that the proposed algorithm comprises 1.21× parameters, with an average intersection ratio of 71.4%. At the same time, the network model and algorithm proposed are contrasted with other equally sophisticated techniques. The comparison findings demonstrate that our approach successfully enhances the accuracy and real-time performance of lane image segmentation in comparison to alternative methods.