Real time alerts on existing CCTV: there are five paths and four of them quietly assume you have IP cameras

Real time means the alert lands on a phone while the operator can still take an action that changes the outcome. That is usually single-digit seconds from the frame being captured to the SMS or chat thread buzzing. If the alert lands after the person has already left the property, that is a faster index into recorded footage, not a real time alert. The cutoff is the action window, not the wall-clock latency in milliseconds.

They do fire in real time, but they are pixel-difference math against a drawn rectangle, not detections of a person. They alert on raccoons, headlights, rain on the lens, sunset gradients, spider webs in front of an IR illuminator, and the apartment dog walking back from the dumpster. Most teams stop checking the alerts within two weeks because the false positive rate is over 95 percent. The recorder is technically alerting in real time. The operator is not getting useful information.

Only if your cameras are IP and you can pull a usable RTSP stream off them. Frigate runs detection on substreams pulled from each camera, then writes events to its own NVR. If your cameras are analog HD-TVI, HD-CVI, or AHD over coax (the dominant install for properties built before 2018), there is no RTSP to pull. If your cameras are IP but the manufacturer has locked the RTSP path or rotated credentials behind a vendor cloud, the same problem appears. Frigate is a great fit for a fresh IP-camera build. It is not a fit for arbitrary legacy CCTV.

Every DVR and NVR has an HDMI output. That output drives the wall display in the management office: a composited multi-view of every camera the recorder knows about. An HDMI capture box plugs into that output, splits the multi-view back into per-tile streams, and runs detection per tile on a local accelerator. The capture is downstream of every camera and recorder decision, so it does not care whether the cameras are analog coax or IP, whether the recorder supports RTSP, or whether the manufacturer has locked the firmware. If the recorder can drive a wall display, capture works. The tradeoff is detection quality scales with per-tile resolution after the crop, so a 16-tile view on a 1080p output gives each camera roughly 480 by 270 pixels of working area.

On HDMI capture with on-prem inference, the perceived latency from intrusion to phone notification is typically 3 to 8 seconds. The hops are HDMI capture (sub-second), per-tile inference on the edge accelerator (100 to 400 ms), event de-duplication and threat classification (sub-second), thumbnail crop (sub-second), push to the messaging provider (1 to 3 seconds), provider fan-out to the device (1 to 3 seconds). On a cloud upload path, an uplink round trip stacks on top: the camera or recorder uploads each frame or each clip to a cloud service, the service runs inference, then the alert flows out. On a typical multifamily DSL or business cable line the uplink is the unstable hop and adds 2 to 30 seconds depending on the upload bandwidth and cloud queueing. The variance is the problem, not the average.

AI-based paths (Frigate, camera-side AI, HDMI capture with an edge AI box, cloud AI services) all reduce false positives by replacing pixel motion with track-based person detection, so insects on the lens, headlights, animals, and rain stop generating alerts. Built-in DVR motion alerts do not. The remaining false positives on AI paths are humans doing legitimate things (a vendor at 3 AM, a tenant returning home through the garage at 1 AM). Those are not detection problems, they are property-context problems, and they need rules for known schedules or known vehicles to suppress.

Depends on the inference budget and the multi-view layout. Cyrano's current edge box supports up to 25 camera feeds per unit. The recorder's HDMI output is configured to display a 4 by 4 or 5 by 5 grid (depending on how many cameras the property has), the capture box reads that grid, and the on-device accelerator runs per-tile inference at a few frames per second per tile. Per-tile resolution drops as the grid grows: a 5 by 5 grid on 1080p output gives each tile about 384 by 216 pixels, which is enough for person detection but starts to limit fine-grained classifications.

No. Camera replacement is a separate axis from real time alerts. Replacing cameras gives you camera-side AI, higher resolution per feed, and modern wiring. None of those are required for real time alerts. An HDMI capture path runs on whatever the recorder produces, including 720p analog HD-TVI feeds composited onto a 1080p wall display. If you replace cameras later, the capture path benefits from the higher per-tile resolution but does not require the swap. If you have already paid for the cameras and the coax pulls and the recorder, replacing them just to get alerts is a sunk-cost mistake.

That is a legitimate sixth path on modern systems and worth considering before adding any other hardware. The honest tradeoffs are: pricing typically scales per camera per month and gets expensive at portfolio scale, the AI runs in the vendor's cloud (so an internet outage takes the alerts down), the alerts use whatever event taxonomy the vendor has chosen (rarely property-specific), and the data leaves the building (which can be a tenant privacy issue depending on the jurisdiction). On most properties built before 2020 the vendor cloud either does not exist or charges enough per camera per month to make a one-time HDMI capture box plus a flat monthly fee cheaper across a 16 to 25 camera install.