New pipeline detects traffic accidents using vision-language models

Researchers have developed a novel pipeline for detecting traffic accidents in video footage without requiring labeled training data. This system, named ACCIDENT @ CVPR, utilizes a coarse-to-fine approach that combines a frozen Qwen3-VL-32B-Instruct vision-language model with YOLO11x object detection and BoT-SORT tracking. The pipeline first identifies a potential collision moment and then refines the analysis using annotated vehicle identities and bounding box coordinates. This method achieved a harmonic mean score of 0.504 on a real-world CCTV test set, outperforming existing baselines by a significant margin. AI

IMPACT This research demonstrates a novel application of vision-language models for real-world event detection, potentially improving traffic safety analysis.

RANK_REASON Academic paper detailing a new method for traffic accident detection. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• ACCIDENT @ CVPR
• alphaXiv
• arXiv
• BoT-SORT
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• Qwen3-VL-32B-Instruct
• ScienceCast
• YOLO11x

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →