Welcome

Thank you for visiting my profile page.

I am a M.S. candidate in Information and Communication Engineering at Myongji University (MJU), advised by Prof. Jaehee Jung .

My master's research focused on enhancing dynamic obstacle avoidance performance in mobile robot navigation by leveraging deep learning models to process semantic information and generate human-like trajectories based on imitation learning.

My primary research interests are focused on:

Learning-based path planning

Imitation learning

Representation learning

Robot Development and Implementation

All of these projects were conducted solely by myself.

Swerve-Drive Robot

Omni-Wheel-Based Robot

Differential-Drive Robot

Research

(* denotes equal contribution)

1. HADP: Hybrid A*-Diffusion Planner for Robust Navigation in Dynamic Obstacle Environments

Mingyu Kim*, Chanyeong Heo*, Jaehee Jung

IEEE Access (Under review)

We proposed the Hybrid A*-Diffusion Planner (HADP) to overcome the generalization and data-efficiency limitations of pure diffusion-based planning in dynamic environments. Our approach uses A* for global path planning in obstacle-free regions and a conditional diffusion model fed by a semantic map, robot pose, and local goal to generate responsive local avoidance trajectories upon detecting dynamic obstacles.

2. HiMSELF: A Hierarchical Misbehavior Classification with Sequence Embedding by Latent Features in Vehicular Ad-Hoc Networks

Mingyu Kim, Dae Hyun Yum, Jaehee Jung

IEEE Access (Under review)

We propose HiMSELF, a misbehavior classification system that learns latent sequence embeddings of multi-class BSM data, applies hierarchical clustering to construct a two-stage classification hierarchy, and achieves an average F1-score of 0.9918 on 19 misbehavior classes, outperforming existing models.

3. Semantic Information Loss Function: A Novel Approach Addressing the Limitations of Pixel-Based Segmentation Loss in Medical Image Segmentation

Chanyeong Heo, Mingyu Kim, Jaehee Jung

IEEE Access (Under review)

We propose Semantic Information Loss, a novel object-level segmentation loss that enhances spatial and structural sensitivity by labeling and matching individual mask regions. This approach involves extracting their size and location features and penalizing any discrepancies, followed by integration with conventional pixel-based losses.