CA-Seaformer: a road image segmentation algorithm based on improved Seaformer

Road segmentation is crucial for autonomous driving, as it automatically identifies and separates different objects or areas in road scenes—like lanes, pedestrians, and traffic signs—to boost the safety and efficiency of these systems. However, the task is complicated by the complexity of road scenes, which often include diverse elements like lane lines, traffic signs, buildings, pedestrians, and obstacles. This complexity can hinder algorithm performance, especially when road markings are unclear or the scene is highly variable. In this paper, we introduce CA-Seaformer, a road image segmentation algorithm based on an improved Seaformer and Coordinate Attention. Our algorithm balances segmentation accuracy with computational efficiency by using Coordinate Attention to enhance both local and global features, boosting model accuracy and lowering computational overhead. Tests on the CityscapesgtFine dataset reveal that CA-Seaformer surpasses other models, hitting an mIoU of 59.32%, Dice of 72.19%, and Precision of 77.09%, with gains of 21.36%, 22.68%, and 6.82% over Seaformer, respectively. These results demonstrate CA-Seaformer’s strong capabilities in precise localization and detail extraction, making it particularly well-suited for resource-constrained environments and fine-grained segmentation tasks.