# DSC-LLM: Driving Scene Context Representation-Based Trajectory Prediction Framework with Risk Factor Reasoning Using LLMs

## Bibliographic information / 논문 정보

- Authors as listed: Sunghun Kim, Joobin Jin, Seokjun Hong, Dongho Ka, Hakjae Kim\*, Byeongjoon Noh\*
- Archive year: 2025
- Category: International Journal
- Status: Published
- Venue / registration record: Sensors 25(23), 7112, 2025.
- Public lab record: https://vita.hanyang.ac.kr/publications/2025-kmmwsx-dsc-llm-driving-scene-context-representation-based-trajectory-prediction-framework-with-risk-fac/index.html
- Text record: https://vita.hanyang.ac.kr/publication-detail/record-b585a3759136b5b4.txt
- Markdown record: https://vita.hanyang.ac.kr/publication-detail/record-b585a3759136b5b4.md
- Original source: https://doi.org/10.3390/s25237112
- DOI: 10.3390/s25237112

## Research summary / 연구 요약

DSC-LLM은 차량의 움직임과 주행 장면의 의미 정보를 결합해 미래 궤적을 예측하고 관련 위험 요인을 설명하는 프레임워크다. 궤적·속도·방향 등의 시간 정보와 전방 영상의 장면 맥락을 결합한 예측 모듈 뒤에 LLM 기반 설명 모듈을 연결한다. Rank2Tell 자료에서 궤적 예측 오차와 구성요소의 효과를 평가하고, 여러 모델이 생성한 위험 설명을 정성적으로 비교했다. 설명 모듈의 평가는 보정된 위험 확률이나 실제 사고 감소를 검증한 결과가 아니며, 예측 모듈의 속도를 전체 시스템의 처리 속도로 해석해서는 안 된다.

DSC-LLM combines vehicle motion with semantic driving-scene information to forecast trajectories and describe relevant risk factors. Its modules extract object behavior and scene context, fuse these representations for trajectory prediction, and generate contextual explanations through a language-model component. Experiments on Rank2Tell evaluate trajectory errors and the effects of individual components, alongside qualitative comparisons of generated explanations. The results support combining scene context with motion forecasting, while explanatory plausibility is assessed separately from calibrated risk prediction. The study does not demonstrate collision reduction, and the reported predictor runtime does not measure the complete reasoning pipeline.

Topics: 주행 장면 맥락; 차량 궤적 예측; 멀티모달 AI; LLM 위험 설명; 자율주행

## Research areas / 연구 분야

AI · 인공지능 / Artificial intelligence; 딥러닝 / Deep learning; 교통 안전 / Traffic safety; 자율주행 / Autonomous driving

## Source and scope / 출처와 범위

This is a lab-maintained overview, not an abstract supplied by the publisher. The archive year and author spelling follow the publication list; exact dates and affiliations are not inferred.

출판된 논문 본문의 연구 방법, 평가 결과와 적용 범위를 확인해 작성한 요약입니다.

This summary is grounded in the published article’s methods, evaluation results, and scope.
- [Published article (DOI)](https://doi.org/10.3390/s25237112)

## Related research / 관련 연구

- [Driving Scene Context-Augmented Trajectory Prediction with Risk-Aware Decision Reasoning using Multimodal LLMs](https://vita.hanyang.ac.kr/publication-detail/2025-driving-scene-context-augmented-trajectory-prediction-with-risk-aware-decision-reasoning-u.html)
- [Traffic Context-Augmented Vehicle Trajectory Prediction Framework using Multimodal LLM](https://vita.hanyang.ac.kr/publication-detail/2025-traffic-context-augmented-vehicle-trajectory-prediction-framework-using-multimodal-llm.html)
- [Anticipatory driving maneuver for safety and comfort: Temporal graph cross attention-based front vehicle trajectory prediction](https://vita.hanyang.ac.kr/publication-detail/2026-anticipatory-driving-maneuver-for-safety-and-comfort-temporal-graph-cross-attention-based-.html)
- [Attention-Driven Lane Change Trajectory Prediction with Traffic Context in Urban Environments](https://vita.hanyang.ac.kr/publication-detail/2025-attention-driven-lane-change-trajectory-prediction-with-traffic-context-in-urban-environme.html)