Researchers develop zero-shot AI pipeline for detecting and classifying traffic accidents in surveillance video without labeled training data

Pipeline uses three independent modules: frame-difference analysis for temporal localization, optical flow for spatial impact location, and CLIP embeddings for collision type classification

Submitted for ACCIDENT @ CVPR 2026 challenge requiring prediction of when, where, and what type of traffic accidents occur

Requires no domain-specific fine-tuning, relying entirely on pre-trained model weights including Farneback algorithm for optical flow and CLIP for image-text similarity

Leverages multi-prompt natural language descriptions and cosine similarity matching to classify collision categories without access to real-world labeled accident data