Dahua NVR people counting: what the setup guides skip, and what to do when the menu greys out

The three gates between you and a working count

The setup wizard makes the path look like one screen. It is actually three preconditions, and each one fails on a different shape of property.

Miss one and the menu may still let you click through, but the count is either zero, double, or wildly off. The Dahua field guides treat this as a configuration error you can fix by tweaking the OSD overlay. It is not a configuration error. It is a hardware constraint that determines the answer before the screen even loads.

What the "AI" sticker on a Dahua NVR actually buys you

The product page says "AI". That word is doing a lot of work. The spec sheet for a typical mid-tier AI-by-NVR unit (think NVR5208-XI or an 8-channel 2AI variant) breaks down like this when you read past the marketing line.

Three things to notice. First, on the mid-tier AI-by-NVR units the AI runs on a small fraction of the channels. Plug in 8 cameras and AI covers 1 to 4 of them depending on which analytic. Second, people counting itself is usually delivered through the camera (WizMind or IVS people-counting model) rather than as an NVR-side analytic on these mid-tier boxes; AI-by-NVR typically lists face and perimeter, not people counting. Third, the data the NVR retains for reporting auto-rolls at six months. Year-over-year retail or tenant flow comparison means exporting the data before the rollover.

Higher-end recorders (NVR5216-16P-XI, the N98A series) have larger AI budgets, but they are also several times the cost of a basic NVR4xxx and are not what most multifamily, retail, or jobsite properties have plugged into the wall.

The native setup, and where it falls off

For the channels where it does work, the path through the menu is straightforward. The block at the bottom of this code is the part that catches most operators by surprise.

On a fleet that satisfies the three gates, that is the whole setup. Pick the channel, draw the area to constrain detection (if you skip this step the count includes people walking past in the background of the frame), draw the tripwire perpendicular to the door with the arrow pointing the way that should count as "Enter," toggle the OSD overlay so the live count shows on the wall display, and apply.

On a fleet that does not satisfy the three gates, you will hit the grey-out at one of three places: the People Counting toggle is disabled in Smart Plan because the camera does not run the analytic; the toggle is enabled but the count never increments because the recorder has no AI slot free; or the toggle works and the count runs but reads zero because the camera is mounted at the wrong angle for the analytic to fire. The grey-out is the cheapest of those failures because it tells you up front. The silent zero is the worst because the NVR happily reports a count and you only notice the count is wrong months later when someone audits the foot-traffic export.

The path that doesn't care which camera you have

When the property has a Dahua NVR with a wall display already running in the leasing office, the cheapest way to count people on every camera (not just the WizMind ones) is to read the same signal that wall display is reading. The recorder is already drawing every camera onto a multiview tile grid. A small box plugged into the NVR's HDMI output captures that grid, splits it back into per-tile crops, and runs detection plus counting on each tile.

From the NVR's side, nothing changes. No firmware, no licenses, no AI channels consumed. The cameras still record locally to the recorder's drives the same way they always did. The HDMI box runs detection on its own SoC, on-site, with no cloud round trip required for the count itself. From the operator's side, the tripwire-and-area UX moves from the NVR's local menu to a web dashboard that shows every camera in one place. People counts are a derived product of the detection model: any camera aimed at a door produces an enter/exit count, regardless of generation. A 2014 HDCVI analog camera composited into tile 7 of a 4x4 grid gets the same treatment as a 2024 WizMind in tile 1.

The trade-offs are real and worth naming. A camera shown in a 5x5 multiview tile is downscaled to roughly 384x216 pixels by the time HDMI captures it, which is enough for person counting and basic classification but not enough for license plate reading. The HDMI box does not see ONVIF metadata, so analytics that rely on PTZ position or vendor-specific event streams are out of scope. And the operator has to disable the NVR's auto-cycle and lock the multiview to a fixed grid so each camera occupies a known tile across reboots. None of these matter for counting people at a doorway.

How to decide which path you actually want

The honest version of this decision is not "Dahua native versus a third-party box." It is "how many doors do you need to count, and what else do you need from the same hardware." A few cases:

• One or two doors, no other analytics needed. Spot-replace those cameras with a WizMind people-counting model, plug into your existing NVR, count on those channels. Cleanest, lowest-touch, no new boxes. Cost is roughly $400-700 per camera installed.
• Three or more doors, plus you also want intent-based alerts at night, plus searchable footage.Replacing every camera adds up fast and still leaves you on the NVR's native search, which is timeline-only. HDMI capture covers every camera the NVR shows, regardless of generation, and adds detection-driven alerts and natural-language search. Cyrano sits in this slot at $450 one-time and $200 per month per property.
• Enterprise retail with strict accuracy requirements and a budget for it.Dahua's 3D stereo people-counting cameras are the right answer here, mounted overhead at the door, paired with an N98A series NVR. This is the deployment the official guides are written for. Most multifamily, jobsite, and small commercial properties are not it.
• You only have HDCVI analog cameras and an NVR4xxx-1U-equivalent recorder.Native people counting is not on the table because neither the camera nor the recorder runs the analytic. HDMI capture is the only retrofit that doesn't require touching the cameras.

The mounting recipe most installs miss, and why it matters either way

Whether the count runs on the camera, the NVR, or an external box, the camera angle is the input every detector ends up working from. Get it wrong and no amount of model swapping rescues the number. Three rules from the Dahua field documentation that hold across analytics:

Mounting height.At least 9 feet above the target with a 3.6mm lens. A 2.8mm lens widens the field but flattens the people in the frame, which makes overlap (two people walking side by side) read as one. A varifocal lens lets you compensate for short ceilings, but the rule of thumb is that you want the camera looking down at people's heads, not at their faces.

Tilt angle. 30-45 degrees down from horizontal at the target. A horizontal shot from a building corner looking out across a parking lot is the worst-case scene for any people-counting analytic, native or otherwise: bodies overlap, perspective is shallow, and outdoor scene complexity (trees, shadows, vehicles) causes the IVS rule to fire intermittently.

Tripwire shape. One straight line per direction, perpendicular to the dominant traffic flow. If entry and exit happen along the same line, the line should be drawn perpendicular to the doorway with the arrow pointing in. Diagonal lines and lines that hug the door frame both produce phantom counts because a person crossing at an angle can register on both sides of the line within the same frame.