## TL;DR
AI surveillance sees everything, everywhere, all at once -- tracking objects across hundreds of cameras, detecting anomalies in real-time, and enabling forensic search through weeks of footage in seconds. The technology raises fundamental questions about privacy, bias, and the balance between security and civil liberties.

## Core Explanation
Surveillance AI: (1) Object detection and tracking -- YOLO detects persons/vehicles, DeepSORT/ReID tracks them across cameras. Multi-camera tracking: re-identification model matches person appearance across disjoint camera views; (2) Anomaly detection -- autoencoders learn normal patterns; deviations flagged. Rare event detection: fighting, running, abandoned object, crowd formation; (3) Forensic search -- query by example ("show me this person across all cameras"), by attributes (red shirt, blue jeans), by time range. Face search where legally permitted; (4) Analytics -- retail (heat maps, dwell time, queue detection), traffic (vehicle counting, wrong-way driving), and perimeter security (intrusion detection).

## Detailed Analysis
BriefCam: forensic video search platform. Indexes video metadata, enables sub-second search across weeks of footage ("find all people in red between 2-4 PM"). Avigilon: self-learning video analytics. Learns normal scene patterns; flags unusual activity without predefined rules. Anomaly detection: autoencoder trained on normal videos (empty corridor, orderly pedestrian flow). Reconstruction error spikes for anomalies (person running, crowd gathering). Edge AI: on-camera processing (Hailo, Ambarella, Qualcomm) eliminates need to stream all video to cloud. Only metadata and alert clips transmitted. Regulation: EU AI Act Article 5 bans real-time remote biometric identification in publicly accessible spaces, except for: targeted search for specific crime victims, prevention of imminent terrorist threat, or identification of serious crime suspects. Chinese Skynet: 600M+ cameras, AI-powered facial recognition integrated with national ID database. US patchwork: no federal law; city-level bans (San Francisco 2019, Boston 2020). NIST FRVT found significant demographic bias -- higher false match rates for women and people of color.