Skip to content
View linearmik's full-sized avatar

Block or report linearmik

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
linearmik/README.md

Seonhyung Kim

Perception for autonomous systems: LiDAR, camera, and the code that turns them into decisions.

B.S. in Artificial Intelligence Engineering, Sookmyung Women's University


About

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.


Tech Stack

Languages

Python C++ C

Perception / ML

PyTorch OpenCV NumPy scikit-learn Pandas

Robotics / Simulation

ROS Isaac ROS MORAI CARLA Jetson Raspberry Pi

Infra / Cloud

Linux Docker CUDA Microsoft Azure AWS Git


Publication

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:
    1. 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.
    2. OneCycle diverged more often as the dataset grew.
    3. 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

Projects

Autonomous Driving Perception · Team FROZEN

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

VLM-Based Drone Navigation · Industry Project

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

End-to-End Driving with PilotNet

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.

MuMulYak · Medication Identification

seonhyung021/winter-team5 · Sookmyung Azure Winter School

Identifies a pill from one photo (Azure Custom Vision + OCR) and generates dosage guidance with Azure OpenAI.

SafeStep · Location-Aware Fall Detection

seonhyung021/SafeStep · Open Source Programming

Owned the ResNet18 stairs/ground classifier (99.2% / 92%+) that sets fall risk by location, integrated with MediaPipe Pose.


Coursework

  • 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

Popular repositories Loading

  1. tourable-web tourable-web Public

    HTML

  2. colorme colorme Public

    HTML 1

  3. MuMulYak MuMulYak Public

    Recognize medications from one photo and get medication guidance, powered by Azure Custom Vision + Azure OpenAI + OCR

    Python

  4. Yolov7_ObjectDetection Yolov7_ObjectDetection Public

    Forked from hopeof-Greatmind/Yolov7_ObjectDetection

    opensourceprogramming

    Jupyter Notebook

  5. HiRN HiRN Public

    Forked from smu-ivpl/HiRN

    HiRN: Hierarchical Recurrent Neural Network for Video Super-Resolution (VSR) using Two-Stage Feature Evolution - Official Repository (Applied Soft Computing)

    Python

  6. autonomous-car autonomous-car Public

    End-to-end self-driving RC car on Raspberry Pi: DL lane detection + adaptive front-vehicle following + obstacle avoidance. Foundation for the ACK 2026 rc-car-scheduling-study.

    Jupyter Notebook