AI Model Predicts Violence Before It Happens From Video

Researchers at the University of Alabama trained a model to predict violence from pre-incident video using facial appearance, audio and body motion. Combining all three signals reached 91.21% accuracy and 96.38% ROC-AUC on the held-out test set.

• Model was trained only on footage preceding violent events, not during them
• Accuracy 91.21%, F1 93.65%, ROC-AUC 96.38% with all three cues combined
• Uses facial appearance, audio and pose-estimated motion from XD-Violence dataset
• Up to four seconds of footage analyzed across three processing branches