Leveraging AI in Surveillance for Public Safety Amid Privacy Concerns

The talk opens with the speaker's reason for building a system that watches existing surveillance camera feeds for firearms and alerts police. The motivation came from a 2023 school bus incident in which an attack was averted only because a gun jammed. The approach described combines visual analysis with thermal and radar inputs, fused by a neural network into a single event so that one situation does not generate several separate alerts.

The middle of the talk surveys other public safety uses, including predictive policing, gunshot localization, behavior-based anomaly detection and facial recognition. Several deployments were described as cancelled or pulled after pilots, and false positives were discussed, such as a camera tripod read as a weapon. The speaker also covered bias in how training images are labeled, the possibility of misidentification, data sharing with little legal restriction, and the absence of a federal standard governing these systems.

For governance, the speaker recommended building privacy in from the start, processing data at the edge, limiting how long footage is kept, and getting public consent. Human review of alerts and confidence thresholds before any action were presented as safeguards, along with publishing accuracy and error rates. The practical takeaway is that accuracy claims and legal limits remain open questions that practitioners and the public still need to shape.