Video analytics tools: what they detect and what they cost

Every product in this market says it does motion detection, and the phrase covers at least three completely different pieces of engineering that behave nothing alike at two in the morning in the rain. Comparing individual products before you know which class of tool you need is how people end up with a quote for a system when what they needed was a setting changed.

So this is the market by class. Five of them, in order of how much they ask of you.

The two questions that separate the classes

What does it look at? Either pixels changed between two frames, or an object was recognised in a frame. Everything cheap looks at pixels. Everything that stops the false alerts recognises objects. The gap between those two sentences is the entire market.

Where does the work happen? On the camera, on a box in your building, or in somebody else's data centre. This decides your bandwidth bill, your privacy posture and what breaks when the internet drops.

Class 1: the detection built into your recorder or camera

Hikvision, Dahua, Reolink, Amcrest, UniFi Protect and the recorder in the cupboard all ship with this. It works by comparing consecutive frames, counting how many pixels changed, and firing when the count passes a threshold inside a drawn region. Newer models add a person or vehicle filter, which is a small model running on the camera and is genuinely better than nothing.

What it detects: change. Rain is change. Headlights are change. A branch is change. A spider on the lens at three in the morning is an enormous amount of change, because it is close to the infrared lamp and therefore very bright.

What it costs: nothing, it is included. That is the reason almost everyone starts here and the reason almost everyone ends up muting the notifications, which is the actual failure mode: the camera keeps recording and nobody is watching.

Worth trying first: tighten the detection region, raise the sensitivity threshold, and set a minimum duration if the interface exposes one. If the on camera person filter exists, turn it on. If that gets you to a quiet night, stop reading, you are done and it cost you nothing.

Class 2: self hosted detectors

Frigate, ZoneMinder, MotionEye, and Blue Iris with an external detector wired in. Software you run on your own machine that pulls the RTSP stream, runs an object detection model over frames, and gives you real classes with real bounding boxes.

What it detects: objects, properly. Frigate in particular is very good, with zone handling that many paid products do not match.

What it costs: the licence is free and the rest is not. You need a machine that stays on, and for anything past a couple of cameras you need an accelerator or a GPU, because CPU only inference falls over as camera count rises. Then you need somebody to own it: updates, disks, certificates, the model that stopped loading after a kernel upgrade. The full arithmetic is in open source video analytics and the cost nobody prices in.

Class 3: cloud video management systems

Eagle Eye Networks and similar. These take your existing cameras through a bridge, store the video in the cloud, and put analytics on top of the recording. You are buying retention first and detection second.

What it detects: objects, plus the ability to search recorded footage, which is a real capability the cheaper classes do not have.

What it costs: per camera per month, tiered mostly by how many days of video you keep, usually sold through a reseller. Upload bandwidth is the hidden line: continuous cloud recording for a dozen cameras is a serious amount of sustained upstream, and a site on a business DSL line should test that before signing anything.

Class 4: analytics layered onto the cameras you own

Camio, Motiondetection, and a handful of smaller products. No new camera, no video storage, no appliance. The software takes the stream or the clip, detects motion, classifies what moved, and hands back an event timeline: what class, when it started, how long it ran, how confident the model was.

What it detects: object classes, filtered by zone and by duration, which is what actually removes the alert storm. What it deliberately does not detect, in our case, is who a person is. There is no face recognition in this product and there is not going to be, which is the sentence that gets a purchase past legal in several US states.

What it costs: in the region of ten dollars per camera per month. Ours is published on the pricing page: $49, $149 and $499 a month for 4, 15 and 60 cameras, which works out at $12.25, $9.93 and $8.32 a camera. There is no free tier; the demo is the free part and it holds nothing back.

Class 5: enterprise platforms

Verkada, Rhombus, Spot AI, Avigilon. Cameras, storage, analytics, access control and a support contract, sold as one system after a call.

What it detects: everything in class 4, plus licence plates, people counting, tailgating at doors, and in most cases face recognition.

What it costs: not published. The pattern is a per camera licence on a multi year term, and in the two hardware cases new cameras as a separate capital line. The number that circulates for the licence sits an order of magnitude above class 4, and the hardware sits on top of that. For a large multi site organisation with a security team this can be entirely the right purchase. For a fifteen camera site it is a very expensive way to fix a software problem.

The cost table at 15 cameras

Scroll the table sideways to read every column. On a phone it is stacked one class at a time instead.

The row that surprises people is class 2. A licence of zero and a hardware bill of a few hundred dollars looks unbeatable until you price the hours, and the hours do not stop after setup.

Which class you are in

• Four cameras, technical, enjoy the tinkering. Class 1 tuned properly, then class 2 if that is not enough.
• Four to forty cameras, nobody on staff to run a server. Class 4. This is the group the product on this site is built for, and it is why every price is published rather than quoted.
• You need the video itself kept for ninety days. Class 3, and put the upload bandwidth in the business case before anything else.
• Fifty sites, a security team, an access control programme. Class 5, and negotiate the renewal terms harder than the first year.

Whichever class you land in, the test is the same. Take a clip from your worst camera at the worst time of night and run it through whatever you are considering. If the tool cannot tell you that the thing at 02:14 was a person and the thing at 02:41 was rain, the class is wrong, not the settings.