TransitNet: A lightweight semantic segmentation network for urban traffic scene understanding.

Existing semantic segmentation networks often suffer from large parameter sizes and high computational complexity, making it difficult to deploy them on resource-constrained in-vehicle systems or edge devices. Additionally, these networks lack sufficient cross-domain adaptability, limiting their performance across diverse scenarios such as traffic and remote sensing. To address these issues, this paper proposes a lightweight semantic segmentation network, TransitNet, with enhanced Cross-Domain Adaptation capability. Based on the PSPNet architecture, TransitNet introduces a Rectangular Context Calibration Attention (RCCA) module to adaptively model the long-range spatial dependencies of road structures and traffic flow distributions. It also incorporates a Bidirectional Fusion Attention (BFA) module to enhance hierarchical feature interaction, thereby preserving fine-grained details. Furthermore, StarNet is adopted as the novel backbone network, leveraging star-shaped operations and depthwise separable convolutions to reduce model parameters. By improving the forward propagation mechanism, it outputs low-level spatial features and upsampled high-level features to support multi-scale feature fusion. Experiments demonstrate that TransitNet achieves an mIoU of 86.98% on the VOC2012 dataset, outperforming PSPNet by 1.58%. On the LoveAD remote sensing dataset, it achieves an mIoU of 61.04%, surpassing CM-UNet by 8.87%. Additionally, TransitNet achieves an mIoU of 76.11% on the OST300 dataset, further verifying its strong generalization and Cross-Domain Adaptation capabilities. By balancing efficiency and accuracy, TransitNet provides a high-performance semantic segmentation solution for real-time environmental perception in autonomous driving systems. It also holds significant potential for applications in fine classification of roads and buildings in remote sensing imagery. The code is available at https://github.com/Eric-863/TransitNet.