RTSP camera AI detection: how it works, and where per-camera pipelines break down

No. Any IP camera that publishes an RTSP stream works, regardless of brand or age. AI detection is software that reads the stream the camera already produces; it does not run on the camera itself. On-camera AI (some Hikvision and Dahua models ship with built-in person and vehicle detection) is a separate and more limited thing: it runs on the camera's own small chip and you cannot change the model or the rules. RTSP-based detection runs on hardware you control, so you pick the model, the zones, and the alert logic.

The substream. Object detectors downscale every frame to a small square tensor before they run, typically somewhere between 224x224 and 640x640 pixels, so a 4K main stream just costs decode time without improving accuracy. Detect on the substream (usually 640x480 or 720p) and reserve the full-resolution main stream for recording, where the detail actually matters as evidence. In Frigate this is the detect role versus the record role; in Blue Iris it is leaving "use main stream if available" unchecked. Detecting on the main stream is the single most common reason a setup that should handle a dozen cameras chokes on three.

It depends on substream resolution, detector hardware, and detect frame rate, not a fixed number. A single Google Coral accelerator running 5fps substreams comfortably covers a handful of cameras; pushing past a dozen usually means a GPU or splitting the load across boxes. The cost is per camera because each stream is its own decode-and-inference pipeline. This linear scaling is fine for a house and becomes the planning constraint once you are covering a property with 16 to 25 or more cameras.

Then a per-camera RTSP tool cannot connect, because it has to authenticate to each stream individually. Your options are to factory-reset the camera or NVR (which you may also lack the password to), get the original installer back on site, or use an ingestion path that does not need per-camera credentials at all, such as reading the recorder's HDMI composite output. On commercial properties with installs more than five years old, missing credentials are common enough that they decide the architecture.

Yes. An NVR re-publishes each connected camera on its own RTSP channel path. On Hikvision NVRs, channel 1 is /Streaming/Channels/101 (main) and /102 (sub), channel 2 is /201 and /202, and so on. On Dahua NVRs you change the channel number in cam/realmonitor?channel=2&subtype=0. Point the detector at the NVR's IP address and the right channel path instead of at the camera directly. The substream-versus-main-stream rule still applies to each channel.

It adds a small amount to the alert, not to recording or live view. The detector reads a copy of the RTSP stream; the camera and the recorder are untouched and keep doing exactly what they did before. End-to-end alert latency is mostly decode time plus one inference pass, typically well under a couple of seconds on local hardware. Cloud inference adds the network round trip on top, which is one more reason security setups tend to keep the model local.

Because a notification only matters if someone acts on it. On a home, you are the responder: the alert reaches your phone and you decide what to do. On a property with nobody watching at 2 AM, an alert that lands in an unwatched app changes nothing on the ground. Detection is the easy half of the problem now that good models are free and run on cheap hardware. The half that actually protects a property is routing the detection to a person who shows up, which is why the last step of any setup matters more than the model.