ITSC 2025 Paper Abstract

Paper FR-EA-T38.4

Zhang, Zhenghan (Tongji University), Ye, Hongwei (Tongji University), Li, Yiyang (Tongji University), Wu, Haoyang (Tongji University), Guo, Yuntao (Tongji University), Li, Xinghua (Tongji University)

PLSJ-Framework: Prompt-In-The-Loop Traffic Safety Judgement Framework

Scheduled for presentation during the Regular Session "S38b-Towards Scalable and Trustworthy AI in Connected Mobility" (FR-EA-T38), Friday, November 21, 2025, 14:30−14:50, Coolangata 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Keywords Deep Learning for Scene Understanding and Semantic Segmentation in Autonomous Vehicles, Real-time Cargo Tracking and Intelligent Supply Chain Management, AI, Machine Learning and Predictive Analytics for Traffic Incident Detection and Management

Abstract

With the widespread deployment of traffic monitoring systems, the demand for traffic video parsing has increased rapidly, especially for accident analysis and responsibility determination. However, existing methods face three major challenges: limited scalability for long videos, coarse understanding of dynamic interactions, and insufficient modeling of relationships for judgmental analysis. In this paper, we propose Prompt-in-the Loop Traffic Safety Judgment Framework (PLSJ-Framework), a MLLM-based traffic video understanding framework, to realize modeling and semantic reduction of high-risk interactions. Specifically, the framework consists of three core modules: the keyframe perception and compression module employs a lightweight strategy to extract high-risk fragments, preserving only the core and adjacent frames to improve processing efficiency. The interaction guidance modeling module constructs multiple role behavior hypotheses based on structured cues, and leverages the large model as a causal agent to infer the interaction logic iteratively. Based on inferred results, the responsibility generation module builds a decision template from contextual information and outputs structured analysis text with temporal clues and primary responsibility attribution. Experiments on real traffic video datasets show that our method exceeds the existing frameworks in judgment accuracy and semantic alignment, and is capable of outputting responsibility analysis conclusions that conform to the logic of manual decision-making, with good interpretability and deployment value.