Research

Multimodal-LLM based framework for generating rare traffic scenarios

Build a multimodal framework that generates realistic and risk-aware traffic scenarios.
Employ LLMs to convert prompts or image captions into textual traffic context, which is then transformed into structured driving scenarios.
Train a Text2Scenario generation model to simulate various interactions between vehicles, including abnormal or rare driving behaviors.
Support downstream applications such as abnormal vehicle behavior prediction, traffic flow disturbance analysis, and comparison with conventional traffic simulation methods.
Enable dynamic and context-sensitive scenario generation for autonomous driving development, policy testing, and safety evaluation through user-driven prompt engineering.

Comfort-aware safety improvement through predictive maneuvering in autonomous vehicles

Enhance passenger comfort and psychological safety in autonomous vehicles by enabling early prediction of surrounding vehicles’ maneuvers such as lane changes or cutting-in.
Support anticipatory decision-making by the ego vehicle through monitoring and forecasting of frontal and side vehicle behaviors in real-time traffic flow.
Focus on maintaining smooth and non-disruptive vehicle operation, especially in multi-lane traffic environments where sudden adjacent vehicle actions may compromise comfort.
Contribute to human-centric autonomous driving systems that prioritize not only collision avoidance but also the emotional experience of passengers.
Demonstrate the importance of comfort-aware risk perception and proactive maneuver selection as integral elements of future ITS design.

Context-aware behavior prediction of disruptive and risky driving vehicles

Investigate deep learning-based prediction models to detect and forecast the trajectories of selfish cutting-in vehicles in off-ramp and merging zones, which often disrupt traffic flow and compromise safety.
Integrate micro-level traffic simulation and real-world traffic video data to capture complex interaction patterns between vehicles in congested scenarios.
Analyze lane-change behavior patterns using road surveillance video to characterize aggressive, non-cooperative, or abnormal maneuvers under various traffic conditions.
Establish a multimodal prediction framework that supports the development of risk-aware, autonomous traffic management systems by identifying potential disruptions in advance.
Contribute to the design of safer, more efficient autonomous driving environments through early recognition and response to high-risk agents in urban road networks.

Design a hybrid OOD segmentation framework that combines anomaly detection (e.g., Mask2Anomaly) and semantic segmentation (e.g., Mask2Former) to achieve state-of-the-art performance.
Generate initial anomaly masks using a flexible anomaly segmentation model and refine them through context-aware score correction guided by a fixed semantic backbone.
Apply meticulous boundary refinement using Laplacian filters and Gaussian blur to enhance object edges and suppress noise.
Achieve robust and precise segmentation results across diverse OOD scenarios through multi-stage refinement and mask quality enhancement.
GitHub: TDB

Federated learning algorithm for efficient city-scale data processing

Develop a city-wide federated learning framework for decentralized and privacy-preserving model training across distributed devices.
Detect collapse or abnormal pedestrian behavior using pose keypoints collected from local environments (e.g., CCTVs, edge devices).
Apply a spatio-temporal Transformer model to effectively capture both spatial structure and temporal dynamics of pedestrian motion.
Ensure model generalization and continuous improvement through round-based local training and global aggregation.
Github: https://github.com/Kim-Byeong-Hun/FedPed-Safe

LLM-based language understanding framework for intelligent transportation systems

Propose a Large Language Model (LLM)-based framework to process and understand diverse natural language data within ITS.
Process legal case documents to predict judicial outcomes such as imprisonment, probation, and fines, highlighting the potential for automated legal reasoning in ITS-related decision-making.
Analyze autonomous vehicle disengagement reports by extracting keywords and classifying causes using LLMs, building a reliable taxonomy of disengagement scenarios.
Structure and interpret unstructured textual data using LLM capabilities, laying the groundwork for integration with Multimodal LLMs (MLLMs) and Vision-Language Models (VLMs) in ITS applications.
Support policy development, legal response, safety assessment, and explainability in autonomous driving systems through enhanced text-based reasoning.

Vision-based vehicle-pedestrian collision risk boundary estimation in C-ITS

Predict potential risks between pedestrians and vehicles by forecasting their future trajectories in real time.
Estimate dynamic risk boundaries based on spatial and temporal patterns extracted from cooperative agents in C-ITS environments.
Apply deep learning models to capture motion intent and interaction behavior between heterogeneous road users.
Enable proactive warning services by identifying imminent conflicts and supporting timely decision-making.
Support real-time, context-aware risk mitigation strategies for enhanced safety in next-generation intelligent transport systems.

Seeing the city: Flows, space, and data

Survey patterns of pedestrian fatalities in relation to urban road environments using spatial analysis.
Analyze the correlation between the deployment of speed enforcement cameras and the occurrence of traffic collisions.
Investigate walkability conditions around elementary schools by incorporating crosswalk delays and pedestrian accessibility using GIS-based walkshed analysis.
Utilize diverse urban-scale datasets to propose data-informed strategies for building safer, more livable neighborhoods.

Page updated

Google Sites

Report abuse