Release Notes#

Version 1.0 (January 2026)#

Initial release of the NVIDIA Public Safety AI Blueprint.

Tailgating Detection – CV-based perception pipeline for detecting unauthorized entry at secure access points
VLM Alert Verification – Cosmos Reason2 8B validates alerts to reduce false positives
AI Agent – Natural language interaction powered by Nemotron Nano 9B v2 for querying incidents and generating reports
Dashboard – Visualize alerts and incident metadata
Kibana Integration – View and analyze security events
MCP Integration – Agent tools for video retrieval, analytics queries, and report generation

NIMs have been tested on H100; configurations may need adjustment on other hardware
VLM analysis of very long videos requires extended processing time
The default LLM has a limited context window; large amounts of incident data may exceed capacity
The RT-CV and Behavior Analytics solution serves as a simple example and therefore operates under specific conditions. An incident is triggered upon detecting at least one individual within the field of view (FOV), and ends once no individuals remain in the FOV for a predefined duration (default: 4 seconds). This method is most effective when incidents are separated by sufficiently long intervals without any person present in the FOV.
Sample videos have been modified with blank frames appended at the end to ensure compatibility with the RT-CV and Behavior Analytics solution.
The Kibana dashboard currently contains no data.
Occasional hallucinations in VLM reasoning have resulted in intermittent inaccurate verdicts.
The report generation agent may, at times, input incorrect values for the “Number of Person Entering” field.