Your IP HD security camera system has one port that makes AI easy, and it is not the one you were sold.

An IP HD security camera system is a set of network-connected cameras (each with its own IP address) recording at HD resolution or higher, feeding into a Network Video Recorder over PoE ethernet. The cameras run 1080p, 4MP, 5MP, 4K, or higher. The NVR stores footage, drives a live wall monitor over HDMI, exposes a web UI or mobile app, and usually supports ONVIF for third-party integrations. The shorthand IP HD covers cameras from Hikvision, Dahua, Lorex, Reolink, Amcrest, Uniview, Swann, Annke, Night Owl, and their OEM rebrands. Where the SERP falls short is everyone sells you the hardware then walks away. Nobody tells the existing-system owner how to make those feeds intelligent without ripping them out.

Because ONVIF and RTSP are per-camera integrations. Every camera has its own credential, its own firmware, its own quirks, its own ONVIF profile support level (S, T, M, or a custom subset). A mixed-brand system, which is most real properties, breaks three ways: a 2019 Hikvision camera behaves differently from a 2023 Hikvision camera on ONVIF, a Reolink refuses RTSP until you toggle a setting behind a menu, and an Amcrest changes its auth scheme after a firmware push. The NVR's HDMI out is downstream of all of that. The NVR has already solved authentication, codec normalization, and multiview rendering for its wall monitor. Whatever you point an HDMI cable at receives one clean 1920x1080 30fps composite frame, every frame, regardless of which cameras are feeding it.

Most consumer and property-grade IP NVRs output 1920x1080 at 30fps over HDMI for multiview, with some 4K HDMI outputs on higher-end units. For tile-scheduled AI the 1080p composite is more than enough. On a 4x4 multiview, each tile is 480x270 pixels, and a person standing 8 meters from a parking-lot camera occupies 60 to 110 pixels of tile height. That is well above the minimum subject size a modern YOLOv8-class detector needs for reliable person-vehicle classification. The compute-architecture gain (one decode, 16 tile polygons, one shared inference loop) outweighs the resolution loss versus pulling 16 native-resolution RTSP streams, because the decode cost at native resolution is the cost you cannot amortize.

It works on every IP HD NVR that has an HDMI monitor output, which is effectively all of them. The supported list includes Hikvision DS series, Dahua XVR and NVR, Lorex LNR and LHD, Reolink NVR, Amcrest NVR, Uniview NVR, Swann, Night Owl, Annke, Q-See rebrands, and Q-See OEMs. The unit does not talk to the cameras. It does not know whether the underlying feeds are 1080p, 4MP, or 4K. It reads the one HDMI signal the NVR produces for its own multiview. Brand, firmware version, ONVIF profile, and per-camera credentials are outside the dependency graph. A property with six Hikvision cameras, four Reolink cameras, and two Amcrest cameras feeding one Hikvision NVR is a single-HDMI integration, not a twelve-way credential juggling problem.

Tile polygons. On a 4x4 multiview layout, the NVR paints the composite as 16 equal rectangles with a gutter and per-tile labels burned in. The Cyrano unit detects the layout within the first second of HDMI signal lock, maps polygons to the 16 tile rectangles, and reads on-screen text labels (parking-north, front-entry, hallway-2) to assign tile-to-camera identity. From that point forward each inference result is tagged with tile id, zone label, and track id. If the operator drills into fullscreen on camera 7, the layout_id changes to a single 1920x1080 tile for that camera, the polygon remaps in one frame, and inference continues against the one active tile. When they exit fullscreen, the 16-tile map returns. No reinitialization, no lost events.

For forensic reconstruction (license plate OCR at 40 meters, facial recognition at distance, evidentiary still capture) the native 4K stream on the NVR's disk is the authoritative source. The NVR preserves that. Cyrano's tile-scheduled compute is a real-time awareness layer on top of the recording layer, optimized for the events a property operator acts on: was someone in the back parking lot for 90 seconds at 2am, did someone tailgate through the front entry, did a package sit in the mailroom past the dwell threshold. For that operational layer, 1080p composite tiles are sufficient. The 4K footage is still on disk for the investigator who arrives the next day.

Adding a cloud AI service to an IP HD system involves per-camera RTSP URLs, per-camera credentials, a site-wide VLAN plan, often a firmware update per camera, and an uplink bandwidth audit. Two to six hours per site with two people is the realistic number, not counting the bandwidth cost of pushing 16 HD streams up a business-class DSL or cable line. Replacing the cameras with smart ones costs $50k to $150k per property in hardware and labor. Plugging a Cyrano unit into the NVR's HDMI out takes under 2 minutes: HDMI splitter, power, network cable, wait for layout detection. The first tile-level event is live under 2 seconds after HDMI signal lock.

One HDMI splitter sits between the NVR and the leasing-office wall monitor. One leg of the splitter continues to the monitor so the operator sees their familiar multiview. The other leg goes into the Cyrano unit. The Cyrano unit takes a PoE or wall-brick power input and an ethernet cable to the property LAN (or a bundled cellular fallback). No camera is touched. No NVR setting is changed. No router rule, no VLAN, no port forward. If the installer needs to leave the property, the split wall feed is still live for the on-site operator.

Per-tile events, each tagged with tile index, zone label, dwell seconds, event class, track id, and a cropped thumbnail. Event classes covered out of the box include person detected in restricted zone, loitering past threshold, tailgating at entry, unauthorized vehicle, package dwell, and zone breach. Alerts route to WhatsApp, SMS, and a dashboard by default. Property teams can route high-threat events to phone calls. Thresholds are set per zone, so the parking-lot tile can have a 60-second loitering threshold while the mailroom has a 10-second package-dwell threshold, on the same unit, sharing the same inference loop.

No. Inference runs locally on the Cyrano unit's silicon, on the same LAN as the NVR. The HDMI frame buffer never leaves the device. When the uplink drops, the unit keeps running person, vehicle, loitering, and tailgating detection, keeps writing events to an on-device index, and queues outbound alerts. When the uplink restores, the alert queue drains. This is a meaningful gap from cloud AI retrofits, which go black the moment the uplink drops, even though the cameras and NVR are still recording locally. Edge compute on the HDMI side keeps the real-time awareness layer alive through outages.

$450 one-time hardware, $200 per month software, per site. On a 25-camera multifamily property with an existing $15k to $30k IP HD system already installed, that is roughly 1.5% to 3% of the deployed hardware cost to add real-time AI across every camera. On a per-camera basis at 25 tiles, the math works out to $18 per tile in hardware and about $8 per tile per month in software. A retrofit cloud-AI-per-camera service typically charges $10 to $30 per camera per month and requires the bandwidth and credential work above before you get a single event.

Three matter. First, if the NVR's HDMI output is broken or disconnected, the unit has no input. Most properties have an active wall monitor driven by that HDMI out, so this is a known-good signal in practice. Second, if the NVR's multiview layout is non-standard (a custom 1+12 or 1+5 mosaic), layout detection uses the fallback manual polygon draw in the install UI. Third, if a camera in the multiview has its on-screen label disabled or the label is purely iconographic, the installer assigns the tile-to-camera mapping manually. All three are bounded install-time issues, not production runtime risks, and all three are resolved once during setup.