AI for Video Surveillance: A Job Description, Not a Camera Spec

1. The job description

If you posted the job on a job board, it would read like this. Watch a 16 to 25 camera multiview on a monitor. Notice when something on the shortlist happens. Within a few seconds, text the property manager with a thumbnail, the camera label, and a one-sentence description. Do not call unless the shortlist event is one of two or three that require escalation. Do this continuously, without fatigue, without bathroom breaks, without lapses after hour three of a shift.

Once you write the job that way, two things become obvious. The sensor on the camera is not the interesting question; the attention model is. And the output surface is not a dashboard; it is a message thread.

2. The six-detection shortlist

A model that recognizes 300 object classes but alerts on every car in the lot is worse than no AI at all. The detection list that actually moves incident numbers at residential and commercial properties is short:

• Person in a restricted zone after configured hours (pool, rooftop, package room, leasing office).
• Loitering past a configured dwell time in entries, stairwells, mailrooms.
• Tailgating through a gated entry or garage arm.
• Package left unattended past a threshold, or someone handling a package they did not deliver.
• Vehicle parked in a fire lane or tow zone.
• Crowd formation at an entry or common area.

Six detections, well tuned, catch most of what on-site staff actually need to respond to. Any AI for video surveillance setup that prioritizes class count over tuning these six is optimizing for a number the property does not care about.

3. Where the AI actually lives

Three real architectures exist. On-sensor inference in a new AI-native IP camera. Server-side inference pulling RTSP or ONVIF streams from each camera. And an edge device that taps the DVR's HDMI multiview output and runs inference on the same grid a guard would have watched.

For new construction, any of the three can work. For a property that already has 16 to 32 cameras wired, only the third is trivially deployable. A single edge device handles up to 25 tiles from one HDMI output, which lines up almost exactly with the 16, 24, and 32 camera counts typical at small, mid, and large multifamily and commercial properties. That number matters because it sets the unit of deployment: one DVR head, one edge device, no server rack, no second network.

The install is five steps on site: unplug the monitor cable from the DVR, plug the edge device's HDMI input into the DVR, plug its passthrough into the monitor, put the device on the network, and label each tile with the camera's real name. The guard station sees the exact same screen it saw before. The AI now sees it too.

4. Delivery: one WhatsApp thread per property

This is the part most guides miss. AI for video surveillance has to deliver its output to a surface the responder already uses. For property managers, maintenance leads, and on-call staff, that surface is WhatsApp or SMS. Not a VMS dashboard. Not an email digest. Not a mobile app they have to remember to open.

The pattern that works: one thread per property. Each alert arrives with a thumbnail, a short clip link, the camera label, the detection class, and the timestamp. The responder replies to mute, forward to dispatch, or acknowledge. The thread itself becomes the incident log. Auditing a month is scrolling a chat, not querying a database.

End-to-end latency from a person stepping onto a restricted rooftop at 2am to the alert landing on a phone is typically a few seconds. The bottleneck is WhatsApp delivery, not the inference.

5. How to tune it without training staff to ignore alerts

Every AI video surveillance deployment fails the same way: it ships with too many detections on, the first week is noisy, staff mute the thread, and by week three the system is dark even though it is technically running. The fix is not a better model. It is tile-level tuning before go-live.

• Per tile, set the hours each detection is active. Pool at 10pm to 6am, leasing office never after hours.
• Per tile, set the dwell time for loitering. A mailroom threshold is different from a stairwell threshold.
• Per tile, pick which detections escalate (call) versus which just message.
• Per tile, draw zones when the camera covers both public and private space (a lobby that also sees the street).

This takes a couple of hours at commissioning. Skipping it is how AI for video surveillance projects quietly die.

6. What to actually buy

Match the buy to the state of the property.

• Existing cameras, wired and mostly working. Buy an HDMI edge device that taps the DVR output and delivers alerts to WhatsApp. Keep the cameras, keep the DVR, change the attention layer.
• No cameras yet, new construction. AI-native IP cameras with on-sensor inference are fine, but still make sure the alert surface is messaging and the detection list is short. The hardware is easy. The operational discipline is hard.
• Mixed portfolio across many properties. Standardize on the alert layer first, not the camera layer. If every property sends to the same WhatsApp format, the operations team can scale. If every property has a different VMS dashboard, they cannot.

7. FAQ