CFA-DeepLabV3+: Cross-level Fusion and Attention Network for Lightweight Road Segmentation

With the rapid advancement of robotics, precise environmental perception is essential for tasks such as autonomous driving and outdoor patrols. Road segmentation provides pixel-level semantic information critical for robot navigation. However, existing algorithms face two major challenges: limited datasets for diverse scenarios and high model complexity, which hinder deployment on resource-constrained platforms. To address these issues, this paper proposes CFA-DeepLabV3+ (Cross-level Fusion and Attention DeepLabV3+), a lightweight road segmentation network. It adopts MobileNetV2 as the backbone to reduce parameters and computational cost, and introduces three complementary modules: an Enhanced ASPP (E-ASPP) for multi-scale context modeling, an Adaptive Fusion Attention Module (AFAM) to dynamically balance channel and spatial attention, and a Cross-level Feature Enhancement Module (CFEM) to fuse shallow details with deep semantics. The IDD dataset is augmented with indoor corridors, forest roads, and woodland trails to boost robustness and generalization. Experimental results demonstrate that CFA-DeepLabV3 + achieves 69.40% mIoU, outperforming state-of-the-art lightweight networks while maintaining low computational overhead, offering superior real-time performance and adaptability for mobile robots in complex environments.