A Multimodal Perception Unmanned System Based on GNN and SLAM

The complex environments encountered in emergency rescue scenarios pose severe challenges to the perception and navigation capabilities of unmanned ground systems. This paper designs and implements a ground unmanned rescue system named "SkyEye" based on multimodal perception fusion. The system adopts a four-wheel independent steering and four-wheel drive (4WIS-4WID) omnidirectional ground chassis and integrates a 4D-FMCW LiDAR, structured-light binocular cameras, a high-precision IMU, and Beidou RTK. Through a tightly coupled graph-optimization SLAM framework and GNN-based dynamic obstacle prediction, it achieves robust perception and autonomous navigation in complex rescue scenarios. Experiments show that under clear open-field conditions, the system achieves centimeterlevel positioning accuracy and obstacle recognition accuracy ≥95%. In extreme dense smoke, recognition accuracy remains at 82.6% with stable navigation. The obstacle avoidance response latency is ≤0.5 s in narrow spaces and GNSS-denied tunnels. This paper elaborates the platform kinematic model, multi-sensor spatio-temporal calibration, detailed SLAM implementation, and complete GNN trajectory prediction pipeline, providing a theoretical basis and engineering reference for ground emergency rescue unmanned systems.