Raphael (Suk Min) Hwang

A real-time LiDAR Gaussian Splatting SLAM system that couples fast G-ICP tracking with spherical rasterization-based mapping, reusing tracking covariances to drive range-aware Gaussian initialization, surface normal estimation, and complexity-based pruning/densification. It achieves 86.78% F-score on Newer College at >20 FPS with purely online trajectories.

Developed a LiDAR-camera sensor fusion pipeline for cone detection and color classification on an autonomous driving platform for 2025 University Mobility Competition Korea. LiDAR points are clustered via DBSCAN to extract cone centroids, then projected onto YOLO segmentation masks for color classification, improving classification accuracy by 30% over the earlier camera-first approach. Paired with a curvature-adaptive Stanley controller on an ERP42 platform for full closed-loop autonomous driving.

Built a 1/10-scale autonomous race car with team SKKUDERIA_SKKAI for our first IFAC Roboracer 2026 appearance. Assembled the platform from a 2D SOSLAB GL5 LiDAR, a Traxxas Slash chassis, and an NVIDIA Jetson Orin Nano, and implemented the full perception–planning–control stack.

2024 Winter Undergraduate Research Program project building a real-time lane keeping system for low-light driving on an Agilex Ranger Mini and NVIDIA Jetson Orin. Camera frames are enhanced on the GPU by a pretrained low-light image enhancement model (Zero-DCE++) and passed to a lane detection model (SRLane), fine-tuned to stop misclassifying sidewalks as lane boundaries. Since the control computer was CPU-only, Zero-DCE++ inference was offloaded to the Jetson Orin over a wired Ethernet ROS network, lifting throughput from 0.3 Hz to 10 Hz, and the system kept the robot centered within its lane through outdoor night tests.

Built as the shared simulation platform for an autonomous driving hackathon hosted for HEVEN club, reconstructing the SKKU campus in 3D so six teams could build and test their own driving algorithms on a scale replica of campus roads. Google Earth imagery is processed through COLMAP for structure-from-motion and 3D Gaussian Splatting for scene reconstruction, then converted to a mesh via GauStudio. The resulting geometry was used to build a road network in RoadRunner and integrated into CARLA as the competition environment.