Lightweight Visual Measurement of Tunnel Scenes Based on <scp>SAE</scp> ‐ <scp>DeepLabV3</scp> +

ABSTRACT Scene understanding is a fundamental prerequisite for autonomous robotic navigation and decision‐making, particularly in hazardous environments such as tunnel construction and cable laying. The complex and unstructured nature of tunnel scenes poses significant challenges to human operators, making accurate scene comprehension essential. In this work, we propose a lightweight semantic segmentation network, SAE‐DeepLabV3+, tailored for tunnel environments. The model introduces Shuffle Attention following the four branches of the ASPP module to enhance global context aggregation and enable adaptive focus on critical features during multi‐scale representation learning. To ensure efficiency, MobileNetV2 is employed as the backbone to significantly reduce model parameters and memory consumption. Furthermore, we incorporate an improved Efficient Channel and Spatial Attention (ECSA) module in the decoder to boost feature refinement with minimal computational overhead. Extensive experiments on a self‐constructed tunnel dataset demonstrate the effectiveness of our approach. SAE‐DeepLabV3+ achieves a mean Intersection over Union (mIoU) of 83.67% and precision of 92.33%, outperforming existing methods such as Fcn, LR‐ASPP, PspNet, Unet, and SegFormer. Compared to the baseline DeepLabV3+, our model achieves improvements of +4.67% mIoU and +5.63% precision, while the model has only 5.81 M parameters. The number of FLOPS is reduced by 113.98 G. The model achieves an FPS of 109.88 in the testing environment and 52.8 FPS in the PyQt‐based interface. These results highlight the proposed model's strong balance between accuracy and efficiency, offering a practical solution for safe and reliable tunnel scene understanding in real‐world applications. This makes it a highly suitable core perceptual component for autonomous systems operating in complex and safety‐critical tunnel environments.