Overview#
A camera at a transport hub detects a partially obscured object that the weapon detection model flags at 78% confidence. The frame goes through the four-stage processing pipeline: detection, verification against secondary model output, severity assessment, and automated incident creation with a 10-second pre-buffer clip. The dispatch integration fires an alert to the security team 11 seconds after the object first appeared on camera. Meanwhile, face and licence plate models are running on the same frame stream at the same time, and a fire detection model is active on a separate camera pointed at the fuel storage area. The Surveillance AI domain orchestrates all of this parallel, per-camera AI processing with fault tolerance built in so that a failure in one model does not interrupt the others.
Key Features#
-
Multi-Model Detection: Process video frames through specialised AI models for object detection, weapon detection, fire and smoke detection, action recognition, face detection, licence plate recognition, and pose estimation simultaneously.
-
Four-Stage Processing Pipeline: Events flow through a structured pipeline of detection, verification, severity assessment, and incident creation to reduce false positives and ensure appropriate response.
-
Event Correlation: Correlate individual detections into meaningful events with deduplication and cooldown periods to prevent alert fatigue from repeated detections of the same activity.
-
Automated Incident Creation: Automatically create incident records for verified high-severity events, including video clip extraction with pre and post buffers for complete context.
-
Dispatch Integration: Route critical events to dispatch systems for emergency response, enabling real-time alerting on detected threats such as weapons, fires, or violent activity.
-
Per-Camera Configuration: Configure which AI models and detection profiles are active on each camera, allowing detection to be tailored to the specific monitoring needs of each location.
-
Resilient Processing: Built-in fault tolerance with automatic recovery ensures the detection pipeline continues operating even when individual AI models experience temporary issues.
Use Cases#
Real-time AI video analytics are most impactful in high-footfall, high-risk, or critical infrastructure environments. Key industries include public safety and law enforcement, transport and logistics, and critical national infrastructure.
-
Threat Detection: Automatically detect weapons, fires, and violent activity in real-time video feeds to enable rapid emergency response.
-
Perimeter Security: Monitor facility perimeters for unauthorised access, suspicious objects, and unusual activity with automated alerting.
-
Traffic Monitoring: Detect traffic incidents, vehicle licence plates, and pedestrian activity to support traffic management and law enforcement.
-
Investigation Support: Search recorded video for specific objects, faces, or licence plates to support active investigations with AI-assisted video analysis.
Integration#
The Surveillance AI module connects with the broader surveillance and safety platform:
- Surveillance Platform: Core component of the real-time video intelligence system
- Camera Management: Receives video frames from managed camera fleet
- Incident Management: Automatically creates incidents for critical detections
- Dispatch: Routes critical events for emergency response
- Zone Management: Detection profiles are configured per surveillance zone
Open Standards#
- ONVIF (Open Network Video Interface Forum): Camera discovery and management use ONVIF over SOAP at the device service endpoint, enabling interoperability with cameras from any ONVIF-conformant manufacturer.
- RTSP (RFC 2326 / RFC 7826): Real Time Streaming Protocol is the primary transport for ingesting live video streams from connected cameras into the detection pipeline.
- RTMPS / RTMP: Processed video streams are re-published to a cloud ingest endpoint using RTMPS, the TLS-secured variant of the Real-Time Messaging Protocol.
- ISO/IEC 14496 (MPEG-4 / MP4): Incident evidence clips are extracted and stored as ISO Base Media File Format MP4 containers, with FFmpeg used for clip trimming and WebM preview generation.
- RFC 4648 (Base64): Video frame bytes are Base64-encoded before transmission to AI model inference endpoints, following the standard alphabet defined in RFC 4648.
- ISO 8601 (RFC 3339) timestamps: All pipeline events, incident records, and clip metadata carry ISO 8601 datetime strings, ensuring interoperable time representation across systems.
- GraphQL (GraphQL June 2018 Specification): The AI model registry and event-detection map are exposed for introspection via a typed GraphQL API, with permission checks on every resolver.
- NENA i3 / NG-911: Critical detections such as weapons, fires, and assaults trigger PSAP dispatch via the Argus PSAP integration, aligning with the NENA i3 standard for routing alerts to public safety answering points.
Last Reviewed: 2026-02-24 Last Updated: 2026-04-14