Jihee Han

M.Eng. student in Autonomy and Robotics at UIUC with seven years of experience in autonomous-driving perception at Hyundai Motor Company.

I am a robotics perception and machine learning engineer specializing in computer vision, autonomous driving, 3D geometry, and production-oriented ML systems.

Before beginning my graduate studies, I worked for seven years as a Research Engineer at Hyundai Motor Company, developing and validating camera-based perception systems for autonomous-driving and ADAS applications.

Education

University of Illinois Urbana-Champaign — Champaign, IL
Master of Engineering in Autonomy and Robotics
Aug 2026 – Expected Dec 2027

Relevant coursework: Computer Vision, Applied Machine Learning, Principles of Safe Autonomy

Georgia Institute of Technology — Remote
Online Master’s Coursework in Computer Science
Aug 2025 – May 2026

Relevant coursework: Robotics: AI Techniques

Konkuk University — Seoul, South Korea
Bachelor of Science in Electronics Engineering
Mar 2015 – Aug 2019

Relevant coursework: Artificial Intelligence, Image Processing, Digital Signal Processing

Experience

Hyundai Motor Company — Gyeonggi, South Korea
Research Engineer, Autonomous Driving Perception Technology
Jul 2019 – Jul 2026

  • Developed camera-based perception systems for autonomous-driving and ADAS applications.
  • Worked across object detection, multi-object tracking, camera calibration, 3D object localization, motion prediction, and perception validation.
  • Integrated TensorRT-optimized perception models into ROS2-based front-camera systems and validated them through on-vehicle testing.
  • Built ML data and evaluation pipelines using Python, PyTorch, Apache Airflow, Ray, and Kubernetes.

Selected Projects

Scalable Sim-to-Real Monocular 3D Object Localization

Developed geometry-driven data synthesis and monocular 3D object localization framework that transfers from simulation to real-world driving data.

Designed dual-stream localization network using 2D keypoints, global bounding-box context, visibility masks, and geometric consistency loss to estimate object position and orientation.

Achieved 1.49 m longitudinal mean absolute error within 70 m on real-world driving data using simulation-only training data.

2.5D Object Detection and Keypoint Network

Developed YOLOv5-based vehicle perception model combining object detection, front, rear, and side-view classification, and side-facet keypoint estimation.

The model achieved 0.740 mAP, 0.804 F1 score, and 27.70 ms inference latency on NVIDIA T4.

Converted the model to TensorRT and integrated the detection and tracking pipeline into a ROS2-based front-camera system for on-vehicle testing.

Multi-Object Tracking and Visual Re-Identification

Developed and validated Kalman-filter-based multi-object tracking system, optimizing track management policies and data association logic through sequence-level ground truth and MOTA-based quantitative evaluation.

Integrated visual Re-ID features into multi-object data association, improving MOTA from 33.13 to 40.10 while reducing computational cost by 29.1% through shared detection and Re-ID features.

End-of-Line Camera Calibration

Engineered end-of-line front-camera calibration algorithm estimating yaw, pitch, and roll within 0.0005 degrees of reference values using nonlinear 3D-to-2D projection equations.

Deployed the calibration algorithm for vehicle installation and validation in an L2 ADAS front-camera perception system.

End-to-End Motion Prediction

Integrated Mamba-based interaction block into an end-to-end autonomous-driving motion-prediction model, reducing sequence complexity from O(L^2) to O(L).

Reduced median GPU inference latency from 8.07 ms to 5.60 ms on NVIDIA H100.

Developed CasADi-based trajectory optimizer to generate temporally smooth ground-truth trajectories from frame-wise 3D object labels.

Video Anonymization and ML Data Pipeline

Built active-learning video anonymization and continuous-learning pipeline for privacy-compliant autonomous-driving data processing.

Implemented Apache Airflow workflows and Ray-based distributed inference on Kubernetes and packaged the anonymization engine as an internally deployable Python library.

Improved Dice from 89.11% to 92.61% across new vehicle configurations and unseen driving environments without manual labeling.

ADAS Parking Perception Validation

Built ADAS parking perception evaluation and failure-analysis tools in C#, C++, and Python, validating object detection and 3D localization outputs using camera and LiDAR data.

Projected LiDAR point clouds into the image plane and applied point-in-polygon filtering to isolate object-associated points for localization validation.

Publication

Scalable Sim-to-Real Monocular 3D Object Localization across Diverse Sensor Configurations
IEEE International Conference on Intelligent Transportation Systems (ITSC), 2026

Technical Skills

Programming: C++, Python, MATLAB, C#

Machine Learning: PyTorch, YOLOv5, Faster R-CNN, ONNX, TensorRT, Model Optimization

Computer Vision: Object Detection, Keypoint Detection, Semantic Segmentation, Camera Calibration, Camera Projection, Coordinate Transformations, 3D Object Localization

Robotics and Estimation: ROS2, Multi-Object Tracking, Re-Identification, Kalman Filtering, Data Association, Motion Prediction, Trajectory Optimization

Systems and Tools: Linux, Git, Kubernetes, Apache Airflow, Ray, OpenCV, PCL, CasADi, SymPy, CUDA Profiling

Research Interests

  • Robotics Perception
  • Autonomous Driving
  • Computer Vision
  • 3D Scene Understanding
  • Sim-to-Real Learning
  • Motion Prediction
  • Human-Centered Autonomy

Contact