Anomaly Detection in Surveillance Video — One-Page Reference

What anomaly detection is (learns 'normal', flags deviation — no rule authored) and how it differs from rule-based behavioral analytics; the four model families (reconstruction/prediction autoencoders, one-class, weakly-supervised MIL, and 2025-2026 vision-language methods — model internals owned by the AI section); the honest accuracy story (controlled benchmarks ~96-98% AUC, real-world UCF-Crime ~80-88% AUC, never 100%); the false-alarm/base-rate math (~86,400 scored moments/day, ~10 real anom