Perception for autonomous systems: LiDAR, camera, and the code that turns them into decisions.
B.S. in Artificial Intelligence Engineering, Sookmyung Women's University
I work on the part of an autonomous stack that has to be right before anything else can be: detection, tracking, and filtering of what the sensors actually see. Most of my time goes into LiDAR pipelines, camera–LiDAR fusion, and the evaluation tooling that tells me whether a change helped or just moved the failure somewhere else.
- Perception lead for Team FROZEN, 2026 KASA AI·S/W Mobility Autonomous Driving Competition (MORAI + ROS)
- Undergraduate first author, ACK 2026 (KIPS Fall Conference), on scheduler and data-scale effects in end-to-end driving
- AI/Perception track on an industry project with Aero Robotics Innovation (ARI): VLM-based drone navigation on Isaac ROS
Some of this work lives in private repositories. Happy to walk through it on request.
Languages
Perception / ML
Robotics / Simulation
Infra / Cloud
The Effect of Data Scale and Learning-Rate Scheduling in a Small-Scale Autonomous-Driving Regression CNN: Offline and Closed-Loop Evaluation Seonhyung Kim, Youngmin Park (Sookmyung Women's University) ACK 2026, Korea Information Processing Society Fall Conference · Poster, Nov 6 2026, Sejong University
- 180 training runs: 2 independent datasets (22K, 14K frames) × 4 schedulers × multiple data scales × 5 seeds, plus closed-loop runs on a Raspberry Pi RC car
- Paired t-tests, TOST equivalence testing, Cohen's dz, and a decomposition of the data-scale effect
- Findings:
- Data scale dominates scheduler choice. Scaling up cut validation loss by 41.9–45.6%, while the largest gap between schedulers was 0.4–3.5%. The three conventional schedulers were statistically indistinguishable.
- OneCycle diverged more often as the dataset grew.
- Models with near-identical offline loss still split on closed-loop completion. Offline validation alone does not predict on-vehicle behavior.
- Journal extension in progress: multi-seed closed-loop replication and mechanism analysis
2026 KASA AI·S/W Mobility Autonomous Driving Competition · Perception Lead · MORAI SIM, ROS, Python
Owned the perception stack end to end, from raw LiDAR/camera input to the obstacle list the planner consumes.
- LiDAR detection with camera fusion (YOLO, CLRNet lanes), plus a prefilter stage (HD-map, static/dynamic, persistence) before planning
- Extended detection to 60 m to cover stopping distance at 60 km/h, holding runtime at 6–12 ms (#116)
- Traced barricade ID respawns to objects slipping between VLP-16 rings and reworked the tracker: ID switches −67% (120 → 40), moving-vehicle velocity accuracy 89% → 93% (#107)
- Built the replay and GT-scoring tools used to measure each change
Aero Robotics Innovation (ARI) · AI/Perception · Isaac ROS, ROS 2 · Fall 2026
- Fusing Visual SLAM geometry with ViT/VLM semantics for context-aware drone autonomy (target: Jetson Orin NX)
- Set up the GPU, Docker, and Isaac ROS environment from bare metal; cuVSLAM running on stereo rosbags
seonhyung021/autonomous-car · Basis of the ACK 2026 paper
Camera-only RC car on Raspberry Pi that regresses steering and speed, with a state-machine layer for obstacle avoidance.
seonhyung021/winter-team5 · Sookmyung Azure Winter School
Identifies a pill from one photo (Azure Custom Vision + OCR) and generates dosage guidance with Azure OpenAI.
seonhyung021/SafeStep · Open Source Programming
Owned the ResNet18 stairs/ground classifier (99.2% / 92%+) that sets fall risk by location, integrated with MediaPipe Pose.
- Image Processing: MFC implementations of histogram equalization/stretching, morphology, noise filtering, edge detection (Laplacian, Sobel across YCbCr/RGB/HSI), PSNR, geometric transforms, YCbCr skin-color face detection
- Machine Learning / Pattern Recognition: CNNs, RNNs, LSTM/GRU, attention, backpropagation, regularization, cross-validation
- Applied ML: scikit-learn pipelines (SVM + StandardScaler + RandomizedSearchCV), decision trees, AWS SageMaker Studio
