SAFERNet: Channel, Positional, and Global Attention Fusion for Efficient RGB-T Segmentation in Disaster Robotics

摘要

Real-time RGB and thermal (RGB-T) fusion is vital for disaster robotics, where robots must navigate unstructured, hazardous environments under tight resource constraints. The main challenge is achieving precise scene understanding, especially at object boundaries, while ensuring computational efficiency for embedded deployment. This work proposes a novel Channel, Positional, and Global Attention (CPGA) fusion block for convolutional neural networks (CNNs) that enhances RGB-T fusion by integrating three complementary attention mechanisms: (i) channel attention for enhancing boundary-focused features, (ii) positional attention for capturing long-range spatial dependencies, and (iii) global attention for reducing redundant computation. Furthermore, we introduce SAFERNet (Spectral-Attentive Fusion for Emergency Response Network), a new RGB-T semantic segmentation architecture built on CPGA blocks. SAFERNet fuses RGB and thermal data across five hierarchical levels of dual ResNet backbones in an encoder-decoder configuration optimized for emergency scenarios. The model is benchmarked against state-of-the-art RGB-T segmentation networks using two publicly available urban scene datasets. Efficiency is assessed through real-time performance metrics—FLOPs (floating-point operations), FPS (frames per second), and PARAMS (parameter count)—in relation to segmentation fidelity. To emphasize applicability in disaster response, we further evaluate SAFERNet on a newly annotated subset of 135 images from the UMA-SAR disaster response dataset, featuring eleven semantic classes tailored for search-and-rescue (SAR) missions. An extensive ablation study evaluates the effects of backbone choices, attention modules, decoder setups, and hyperparameters. SAFERNet consistently balances accuracy and efficiency, making it well-suited for real-time deployment on autonomous robots in disaster scenarios. Code and annotated data are available at https://github.com/amsalase/CPGFANet.