Multimodal RGB-LiDAR Fusion for Robust Drivable Area Segmentation and Mapping

Drivable area detection and segmentation are critical tasks for autonomous mobile robots in complex and dynamic environments. RGB-based methods offer rich semantic information but suffer in unstructured environments and under varying lighting, while LiDAR-based models provide precise spatial measurements but often require high-resolution sensors and are sensitive to sparsity. In addition, most fusion-based systems are constrained by fixed sensor setups and demand retraining when hardware configurations change. This paper presents a real-time, modular RGB-LiDAR fusion framework for robust drivable area recognition and mapping. Our method decouples RGB and LiDAR preprocessing to support sensor-agnostic adaptability without retraining, enabling seamless deployment across diverse platforms. By fusing RGB segmentation with LiDAR ground estimation, we generate high-confidence drivable area point clouds, which are incrementally integrated via SLAM into a global drivable area map. The proposed approach was evaluated on the KITTI dataset in terms of intersection over union (IoU), precision, and frames per second (FPS). Experimental results demonstrate that the proposed framework achieves competitive accuracy and the highest inference speed among compared methods, confirming its suitability for real-time autonomous navigation.