VibeNVR v1.28.0 and v1.28.1 Introduce Multi-Model AI Pipeline and Global Master Switch
Enable the global AI master switch to unload models from memory and use MQTT auto‑discovery for Home Assistant integration.
Enable the global AI master switch and configure MQTT auto‑discovery to optimize resource usage and integrate with Home Assistant.
Summary
VibeNVR released v1.28.0 and v1.28.1, adding a multi-model TFLite pipeline that supports YOLOv8 and MobileNet SSD v2, a global AI engine for centralized model and hardware selection, NMS with configurable IoU threshold, hysteresis for smoother tracking, native MQTT service with Home Assistant auto‑discovery, global settings sync, and dashboard health widgets. v1.28.1 adds a universal AI master switch that unloads models from memory, an intelligent OpenCV fallback when the engine is off, a singleton engine architecture to prevent redundant loads, visual state indicators for fallback mode, and RBAC hardening restricting AI configuration endpoints to the Admin role. These updates allow resource‑constrained deployments to unload AI models when not needed, saving CPU and RAM, and integrate VibeNVR events as binary sensors in Home Assistant, simplifying automation rules. The Docker Compose setup requires no cloud or subscriptions, and the full changelog is available on GitHub. The article emphasizes the significance of these features for homelab and constrained hardware, and invites readers to try the new features via Docker Compose.
Key changes
- Multi‑model TFLite support for YOLOv8 and MobileNet SSD v2
- Global AI Engine centralizes model and hardware selection
- NMS with configurable IoU threshold eliminates redundant boxes
- Hysteresis (1.0s persistence) smooths object tracking
- Native MQTT service with Home Assistant auto‑discovery registers sensors automatically
- Global Settings Sync propagates MQTT, AI, and throttling settings in real time
- Dashboard Health Widgets show live MQTT connection and AI processor status
- Universal AI Master Switch unloads models, saves CPU/RAM, with OpenCV fallback and RBAC hardening