Enhancing Small Object Segmentation in RGB-T Semantic Segmentation for Autonomous Driving

With the rapid development of intelligent driving technology, the performance of automotive environment perception systems plays a crucial role in ensuring the safety of autonomous vehicles. RGB-based semantic segmentation struggles in low-light, tunnels, and adverse weather, while thermal infrared imagery offers robustness. However, existing RGB-TIR methods often neglect small object segmentation accuracy, compromising safety. To address the above issues, we propose a novel multimodal semantic segmentation framework integrating RGB and TIR data to enhance small object segmentation. Specifically, we design a Small Object Alignment module for cross-modal feature consistency, and a Small Object Fusion module that effectively integrates both modalities without preferential bias. Experimental results demonstrate a notable improvement in the mean Intersection over Union for small objects in automotive-related scenarios from public RGB-T semantic segmentation datasets.