YOLO와 고도 맵핑의 통합을 통한 지도 구축 및 내비게이션

초록

e present a local mapping and navigation pipeline that fuses robot-centric elevation mapping with YOLO-based object semantics to improve safety and reliability in cluttered environments. Elevation maps are built from either depth-camera or LiDAR point clouds and pre-processed using ROS filters (outlier removal, temporal filtering, ray-based clearing). From the elevation map, we compute a traversability layer (e.g., slope/roughness/step height) and export it via grid_map to a 2-D occupancy/cost map used by the local planner. In parallel, a depth camera feeds a YOLO detector (through yolo_ros) trained on task-relevant classes (e.g., person, bicycle, chair). Detections are projected to the robot frame using calibrated extrinsics between the depth camera and the elevation-mapping sensor; the resulting semantic cues are injected as a dynamic cost layer or velocity constraints (e.g., slow down or stop on human detection). Preliminary experiments on indoor corridors and outdoor walkways demonstrate that the proposed fusion reduces spurious obstacles and yields more stable local paths compared with an elevation-only baseline, while maintaining real-time performance on an embedded CPU/GPU platform. The approach is fully implemented in ROS and requires only standard packages and a one-time extrinsic calibration. The method is applicable to autonomous mobile robots in campuses, factories, and public spaces, and can be extended to additional classes and sensors. Implementation details such as sensor models, filter parameters, and class sets are provided for reproducibility in the final paper.