Most traditional Network Video Recorders are fundamentally broken digital hoarders. They blindly dump raw video streams onto spinning disks, saturating IOPS and forcing operators to scrub through hours of static pixels to locate a two-second security incident. Pushing that continuous stream up to a proprietary cloud provider fixes the detection problem, but it replaces it with recurring subscription bills, bandwidth exhaustion, and severe privacy compromises.

Frigate NVR changes this paradigm completely. Built from the ground up as an open-source, local-first NVR for IP cameras, Frigate executes real-time AI object detection directly on your edge infrastructure. It cuts cloud dependencies out of the equation entirely, keeping your video pipelines fast, private, and deterministic.

The Architecture: Why Your CPU Does Not Melt

Running continuous neural network inference across multiple high-resolution IP camera feeds will instantly saturate even high-end consumer CPUs if implemented naively. Frigate prevents compute starvation by orchestrating a highly optimized, multi-tier processing pipeline.

+------------------+      +-------------------------+      +---------------------------+
| IP Camera Stream | ---> | Low-Overhead Motion     | ---> | Targeted Inference        |
| (RTSP Pull)      |      | Detection (Gatekeeper)  |      | (TensorFlow / Coral TPU)  |
+------------------+      +-------------------------+      +---------------------------+
                                                                       |
                                                                       v
                                                           +---------------------------+
                                                           | MQTT & Event Broker       |
                                                           | (Home Assistant / Alerts) |
                                                           +---------------------------+

1. Motion Detection as an Execution Gatekeeper

Frigate never runs heavy object detection on every raw frame pulled from a camera. Instead, a lightweight motion detection step built on OpenCV acts as an execution gatekeeper. It continuously inspects the frame to determine if pixel deltas justify further compute. If the frame contains no movement, the pipeline immediately drops it before hitting the tensor engine.

2. Region-of-Interest (ROI) Inference

When movement does occur, Frigate does not pass the entire full-resolution canvas into the detector. It isolates the specific bounding box containing the motion, cropping out static backgrounds. Only these precise regions of interest are fed into the model. This eliminates wasted tensor operations on stationary scenery like driveways, walls, or empty lawns.

3. Dedicated Hardware Acceleration and Multiprocessing

Frigate leans into Python's multiprocessing architecture, isolating inference workloads into dedicated, independent processes. This design ensures that raw frame ingest, stream muxing, and AI workloads never block each other, preserving low-latency performance over raw batch queuing.

To push efficiency to its theoretical limits, Frigate strongly recommends offloading inference to dedicated hardware accelerators:

Compute Platform	Role in Pipeline	Operational Impact
Host CPU	Video ingest, motion filtering, orchestration	Baseline processing; vulnerable to compute exhaustion without accelerators
Google Coral TPU	Dedicated edge tensor processing	Ultra-low inference latency with near-zero host CPU load
Intel OpenVINO	Optimized hardware execution	Hardware-accelerated inference across supported silicon

By delegating the matrix math to a Google Coral TPU or an Intel OpenVINO runtime, the host system runs cool while delivering fluid, instant identification.

Network Efficiency and Sub-Second Streaming

Handling multiple concurrent video consumers often degrades camera performance. Most embedded IP camera network stacks buckle under the weight of several parallel RTSP connections.

Frigate mitigates this with an integrated RTSP re-streamer. The host establishes a single upstream RTSP pull from each physical camera, caching and broadcasting that stream internally to downstream clients.

                       +---------------+ ---> WebRTC / MSE UI
[ IP Camera ] =======> |  Frigate NVR  | ---> 24/7 Disk Storage
 (Single RTSP Pull)    | (Re-Streamer) | ---> Home Assistant / Third-Party
                       +---------------+

For live operations, latency is the difference between active monitoring and historic review. Frigate natively integrates WebRTC and Media Source Extensions (MSE), slashing live view latency down to sub-second speeds. Operators can track physical events in real time without the multi-second buffering inherent to legacy HLS streaming protocols.

Storage Policies and Event-Driven Pipelines

Continuous 24/7 video capture is essential for audit trails, but it quickly consumes storage pools. Frigate implements a tiered retention engine that balances raw capture against prioritized events.

• Continuous 24/7 Recording: Maintains an uninterrupted baseline video history on local disks.
• Granular Retention Rules: Allows administrators to establish customized retention intervals by object class. You can configure the system to retain "person" detections for 30 days while dropping "car" events after 7 days, purging unflagged footage automatically.

Event metadata and detection payloads are published immediately over MQTT. This design choice makes Frigate ecosystem-agnostic. While it features a custom component for deep Home Assistant integration, the MQTT backbone allows engineers to pipe structured telemetry straight into Node-RED, bespoke operational dashboards, or custom backend services.

Operator-Grade Management

Tuning spatial detection rules traditionally required manually transcribing coordinate arrays into configuration files. Frigate replaces that friction with an operator-grade web interface:

• Visual Zone and Mask Editors: Draw dynamic detection zones and motion masks directly over live video feeds without touching raw YAML geometry coordinates.
• Event Review and Timeline Scrubbing: Rapidly navigate across multi-camera layouts with interactive timeline scrubbing optimized for fast triage.
• Live Operational Dashboard: Unified pane of glass delivering real-time camera views, hardware health, and instant detection feeds.

Licensing and Governance

Frigate NVR is open-source software distributed under the permissive MIT License. The underlying codebase, configuration schemas, and technical documentation are free to inspect, modify, and deploy.

Note that the "Frigate" name, the "Frigate NVR" brand, and its visual logos are registered trademarks of Frigate, Inc., and are governed by a separate trademark policy outside the scope of the MIT code license.

Project development is sustained directly by the community through GitHub Sponsors, with internationalization coordinated via Weblate. Full deployment guides and architectural documentation are available at https://docs.frigate.video.