Optimizing Your NVR Config: Alert vs Detection Recording, GPU Tuning, and Notification Control

1. Common NVR config mistakes that waste storage and attention

The most frequent mistake with self-hosted NVR setups is recording everything at full resolution, 24/7, on every camera. This fills up drives fast and makes it nearly impossible to find relevant footage when you actually need it. A 16-camera system recording continuously at 2K resolution can burn through 8 to 12 TB per week depending on your codec settings and scene complexity.

The second common mistake is treating all motion events as equally important. Default Frigate configs often start with a broad "detect all objects on all cameras" approach. This floods your event timeline with hundreds of detections per hour: cars on the street, birds, shadows from clouds, and leaves blowing across the frame. When everything is an event, nothing is.

A third issue is not separating your recording retention policies by event type. Frigate supports different retention periods for "alert" recordings versus "detection" recordings versus continuous background recording. If you treat them all the same, you either keep too much (expensive storage) or delete too aggressively (losing footage of real incidents before you review it).

Finally, many users leave notification settings at their defaults. Every detection triggers a push notification, and within a week, the user either mutes the notifications entirely or ignores them out of habit. This defeats the purpose of running object detection in the first place.

2. Alert vs detection recording strategies

Frigate distinguishes between "alerts" and "detections" in its recording configuration. Understanding this distinction is the single most impactful optimization you can make. An alert is typically a tracked object that enters a zone you have explicitly marked as important. A detection is any object that the model identifies, regardless of location.

The recommended approach is to configure your system so that only alerts trigger notifications and get the longest retention. Detections should still be recorded, but with shorter retention and no push notifications. This way, a person walking on the sidewalk in front of your property gets logged as a detection (available for review if needed), while a person entering your driveway or approaching a door triggers an alert with a notification to your phone.

In your Frigate config, this looks like setting record.alerts.retain.days to a longer period (14 to 30 days) and record.detections.retain.days to a shorter window (3 to 7 days). For notifications, configure your MQTT or Home Assistant integration to only fire on alert-level events. Some users also add a minimum score threshold (e.g., 0.7 or higher) before an alert notification sends, which further reduces noise.

For continuous background recording, consider dropping to a lower resolution sub-stream. Most IP cameras support dual streams: a high-resolution main stream for recording events and a lower-resolution sub-stream for continuous background capture. This gives you 24/7 coverage for forensic review without the storage cost of full-resolution continuous recording.

3. GPU utilization and detection resolution tuning

Object detection is the most compute-intensive part of any NVR pipeline. Frigate supports multiple detector backends including Coral TPU, OpenVINO (for Intel iGPUs), and NVIDIA GPUs via TensorRT. Choosing the right backend and tuning your detection resolution has a major impact on both accuracy and system load.

A common misconception is that feeding higher-resolution frames to the detector produces better results. In practice, most object detection models (like the default SSD MobileNet or YOLO variants) resize input frames to a fixed resolution internally, often 320x320 or 640x640 pixels. Sending a 4K frame to a model that downscales it internally wastes GPU cycles on the resize step without improving detection quality.

The optimal approach is to set your detect stream resolution in the Frigate config to match what the model actually uses. For a standard SSD MobileNet model on a Coral TPU, a detect resolution of 1280x720 is more than sufficient. Some users run detection at 640x480 with excellent results, particularly for cameras that cover a relatively small area like a doorway or porch.

For NVIDIA GPU users, TensorRT provides significant performance gains over generic ONNX inference. A single GTX 1060 or RTX 3060 can comfortably handle detection across 8 to 12 cameras when the detection resolution is tuned correctly. Monitor your GPU utilization with nvidia-smi and aim to keep sustained utilization below 70% to leave headroom for burst activity.

If you are running a Coral TPU, keep in mind that USB Coral devices have lower throughput than PCIe variants. A USB Coral can handle roughly 4 to 6 cameras at reasonable frame rates. If you have more cameras than that, consider adding a second Coral, switching to a PCIe model, or offloading some cameras to GPU-based detection.

4. Zone-based detection and reducing false positives

Zones are the most underutilized feature in Frigate. A zone is a polygon you draw on the camera's field of view that defines areas of interest. Objects detected outside your zones can be ignored entirely, logged as low-priority detections, or handled differently than objects inside zones.

For a typical residential setup, you might define a "driveway" zone, a "front door" zone, and a "backyard" zone. A person detected in the front door zone triggers an alert. A person detected on the public sidewalk (outside any zone) gets logged as a detection only. A car detected in the driveway zone triggers an alert, while a car on the street is ignored.

For commercial properties with many cameras, zone configuration becomes even more critical. Parking lot cameras can be zoned to alert only when a person enters an area after business hours. Loading dock cameras can alert when a vehicle approaches outside of scheduled delivery windows. Perimeter cameras can alert on any person detection since nobody should be in those areas.

Beyond spatial zones, Frigate also supports object filters based on size. Setting minimum and maximum object dimensions helps eliminate false positives from small animals, moving shadows, or distant objects on the edge of the frame. For example, requiring a detected "person" to occupy at least 5% of the frame height eliminates most phantom detections from trees, flags, and other swaying objects.

Combining zones with object filters and score thresholds creates a layered filtering system. Each layer removes a category of false positives. The result is a dramatically cleaner event timeline where the events that remain are genuinely worth your attention.

5. The maintenance burden of self-hosted NVR vs managed solutions

Self-hosted NVR platforms offer unmatched flexibility and privacy. Your footage stays on your hardware. You control every parameter. There are no monthly fees for the software itself (though hardware, power, and storage have real ongoing costs). For technically inclined users who enjoy the process of tuning and optimizing, Frigate and Blue Iris are excellent platforms.

However, the maintenance burden is real and ongoing. Frigate config files need updating when you add cameras, change camera positions, adjust detection models, or update the software. YAML config errors can silently break detection on specific cameras without obvious symptoms. Coral TPU firmware updates, Docker container management, storage monitoring, and database maintenance all require periodic attention.

For property managers and business owners who need reliable security monitoring but cannot dedicate engineering time to NVR maintenance, managed solutions are worth considering. These range from cloud NVR platforms like Verkada and Rhombus to edge AI devices that plug into existing camera infrastructure.

One example in the edge AI category is Cyrano, which connects to an existing DVR or NVR via HDMI and runs object detection locally without requiring any config files. The device processes the video feed on-device and sends alerts when it detects relevant activity. This approach works for organizations that already have cameras and a recording system but want to add intelligent detection without replacing hardware or managing software configs.

Neither approach is universally better. If you enjoy the technical process and want maximum control, self-hosted NVR platforms are hard to beat. If you need reliable detection without the config overhead, and you are willing to pay for that convenience, managed or plug-and-play solutions fill the gap. Many property managers start with a self-hosted setup, realize the maintenance burden does not scale across multiple properties, and eventually move to managed solutions for operational simplicity.