GDA-RoadSeg: an improved road segmentation network with gated depthwise attention feature fusion

Road segmentation is an important and challenging task for robotics working in unstructured environments. There are some problems that need to be solved, such as low efficiency in multi-scale feature fusion, insufficient long-range dependency modeling, and inaccuracy of edge segmentation. To address these issues, an efficient road segmentation model based on ResNet34 is proposed in this paper. First, we design a gated depthwise attention fusion module (GDAFM), which dynamically fuses shallow-detail and deep-semantic features via a spatial attention mechanism and depthwise convolution to improve fusion efficiency. Second, we proposed an enhanced asymmetric dilated block (EADB) by employing a large horizontal dilation rate to strengthen long-range dependency modeling and optimizing parameters to eliminate the gridding effect. Additionally, we introduce an edge-aware auxiliary branch (EAB), combining automatically generated edge supervision signals with a multi-task loss function to significantly boost boundary accuracy. Experiments on the Cityscapes and CamVid datasets show that our model achieves MaxF scores of 97.89% and 97.46%, respectively. The results show that our model outperforms other state-of-the-art models.