Release Notes: Scene Understanding Service#

This page tracks releases of the Scene Understanding Service microservice. The most recent release is listed first; older entries are preserved for history.

Version 2026.2.0#

Release Date: September 9, 2026

Scene Understanding Service is a generic microservice for multi-scene, zone-based behavioral analysis. Its behavior is defined entirely through two YAML files (scene-config.yaml and rules.yaml), enabling scenarios such as retail loss prevention, restricted-area monitoring, and other zone-based use cases without code changes.

New:

  • Multi-scene, multi-camera person tracking with per-person session state (zone visits, dwell time, behavioral flags).

  • Declarative rule engine: thresholds and detection rules are changed through configuration, with support for severity escalation and per-session de-duplication of alerts.

  • Optional deeper analysis: escalate selected events to a behavioral-analysis worker (pose + visual language model) for concealment-style detection.

  • Automatic zone discovery from Scenescape at startup, with the ability to add or update zones at runtime.

  • Optional evidence-frame capture and routing of alerts to a downstream alert service.

  • Ships self-contained with sample configuration for standalone evaluation and drop-in use with any Scenescape-based deployment.

Known Issues:

  • The service requires a reachable Scenescape deployment (including the analytics component on Scenescape 2026.2.0) to produce zone-based events and alerts.

  • Alert visibility depends on a reachable alert service.

Version 0.1.0#

First release of the Scene Understanding Service as a self-contained, reusable microservice for multi-scene behavioral analysis and suspicious activity detection, built for edge deployment on Intel hardware.

Release Date: June 18, 2026

New:

  • Scenescape MQTT-driven, multi-scene, multi-camera person tracking with a per-person session state machine (zone visits, dwell time, flags).

  • Declarative YAML rule engine (rules.yaml) producing alert and escalate actions; thresholds and rules change without code edits.

  • Optional behavioral-analysis escalation (pose + VLM) integrated over the ba/requests / ba/results MQTT topics.

  • Zone auto-discovery from the Scenescape REST API at startup, with on-demand re-discovery via POST /api/v1/sus/zones/discover.

  • Optional SeaweedFS evidence-frame capture and alert-service routing.

  • REST API under /api/v1/sus for session, zone, and alert state, plus /health.

  • Self-contained image with bundled sample config (scene-config.yaml, rules.yaml) that runs standalone; consuming applications override via a read-only volume mount.

  • New User Guide doc set: overview, get-started, how-it-works, configuration, api-reference, and troubleshooting, plus an architecture diagram.

Known issues:

  • The service is an event consumer/producer; it requires a reachable Scenescape deployment (MQTT + REST) to produce meaningful output.

  • The alert endpoints depend on a reachable alert-service; they return empty results when alerting is disabled or the alert-service is unavailable.