How do I record RTSP cameras on Linux with object detection and no Coral?
You can record RTSP cameras and detect people, vehicles and animals on a Linux machine with no Coral and no other accelerator. This guide installs Aperture, adds the cameras, turns on object recognition on the CPU and shows what it costs, then shows how an NVIDIA GPU changes the numbers.
The short answer
Section titled “The short answer”- Aperture’s object recognition runs on the CPU or on an NVIDIA GPU. It does not use a Coral. On a CPU, the Balanced model needs about 30 ms per picture on a desktop CPU.
- Recording costs almost nothing in CPU: Aperture copies each camera’s stream to disk without re-encoding it.
- Recognition is off on a new install, and nothing is downloaded until you turn it on.
What you need
Section titled “What you need”- A 64-bit x86 machine running Debian 12 or 13, or Ubuntu 22.04, 24.04 or 26.04, with systemd. A mini PC, an old laptop or a virtual machine works. ARM boards such as a Raspberry Pi are not supported.
- Cameras that serve RTSP, ideally with a second, smaller sub stream. Aperture’s motion detection and recognition read the sub stream when there is one.
- As a reference point, four cameras recording around the clock with motion detection on their sub streams run on a machine with 4 CPU cores and 4 GB of RAM. Add cores for object recognition on the CPU. See Requirements and sizing.
-
Install Aperture. On the machine, run:
Terminal window curl -fsSL https://aperturenvr.com/install | sudo shThe wizard checks the machine, asks a few questions and shows every step and file before it changes anything. Under Options you can choose object recognition on the processor now, or leave it off and turn it on later. See Install.
-
Add each camera. In Aperture, open Devices and click Add camera. Enter a Camera name, paste the camera’s RTSP address into Main stream (RTSP URL) and the sub stream’s address into Sub stream (optional, used for motion). Click Test connection to see a picture, then Add camera. The addresses for Reolink, Tapo, Hikvision, Dahua, Axis and other brands are listed in Add cameras.
-
Turn on recognition. Open Settings › Detection. The card shows where recognition would run (for example “CPU, 3 threads”), which model it would use and roughly how long each picture takes. Select Enable Object Recognition & Download Required Models. When the card shows Running, recognition is active.
-
Try it. In the Try it row, pick a camera and select Try. Recognition runs on a live picture and lists what it found. Live boxes appear in the camera’s live view when something moves.
Recognition looks at cameras that have motion detection on. Motion is always the trigger, so a camera with nothing moving costs almost nothing. See Motion and audio detection.
Choose a model for the CPU
Section titled “Choose a model for the CPU”Automatic picks for your hardware: Fast on a CPU with two threads or fewer, Balanced on any other CPU. These are the models, with the sizes and times Aperture documents (rough figures for a desktop CPU and a mid-range NVIDIA GPU; the settings page shows the time measured on your machine):
| Choice | Model | Input | Download | Per picture, CPU | Per picture, GPU |
|---|---|---|---|---|---|
| Fast | YOLOX-Nano | 416 px | 3.5 MB | about 9 ms | about 2 ms |
| Balanced | YOLOX-Tiny | 416 px | 19 MB | about 30 ms | about 3 ms |
| Accurate | YOLOX-S | 640 px | 34 MB | about 110 ms | about 4 ms |
| Best | YOLOX-M | 640 px | 97 MB | about 300 ms | about 12 ms |
| All-in-one | D-FINE-X (Objects365) | 640 px | 240 MB | about 900 ms | about 24 ms |
The All-in-one model needs an NVIDIA GPU. It is the one that also finds delivered parcels live. On a CPU, a separate small model (D-FINE-S, 40 MB) can watch for parcels instead; see Packages.
A bigger input sees small and distant things better. A very small sub stream limits how far away a person can be recognised.
How much CPU it uses
Section titled “How much CPU it uses”- Aperture uses at most half the machine’s CPU threads, at most half the service’s CPU quota, and never more than four. With the quota the installer sets (six CPUs), that is at most three threads.
- On a CPU, five pictures a second are decoded for recognition. A camera that is moving gets at most one picture to the detector every 300 ms.
- One detector serves all cameras, and they take turns. When the model is slower than the cameras ask for, each camera gets fewer pictures. Nothing piles up.
- If the worker process fails, recording and motion detection carry on, and Aperture restarts the worker.
Recognition shares the service’s CPU and memory limits with recording and the web app. On a small machine, prefer the Fast or Balanced model. See How often it looks and Camera capacity.
Add an NVIDIA GPU later
Section titled “Add an NVIDIA GPU later”An NVIDIA GPU from the Pascal generation (GTX 10 series) to Blackwell (RTX 50 series) makes recognition much faster and allows the larger models. Recognition keeps running on the CPU until the GPU is ready.
- Install the NVIDIA driver. This is the one thing you install yourself. Settings › Detection shows a guide with the exact commands for Debian 12 and 13 and Ubuntu 22.04, 24.04 and 26.04.
- Select Use the GPU & download. The download is about 2.9 GB for the GPU runtime, plus the model. Aperture wants three times the download size free before it unpacks.
- On a GPU, ten pictures a second are decoded, and a moving camera gets a picture to the detector every 100 ms at most.
See CPU or NVIDIA GPU for what the driver needs, including Secure Boot. AMD GPUs are detected but not used; recognition then runs on the CPU.
If you have a Coral
Section titled “If you have a Coral”Aperture’s own recognition does not use a Coral, a Hailo module or any other accelerator besides an NVIDIA GPU. If you already own one, you can run an external detection server that speaks the DeepStack or CodeProject.AI API on that hardware and point Aperture at it under Settings › Detection › Use an external detector instead.
The external detector works differently from local recognition. Aperture sends snapshots when a motion event opens. There are no live boxes, no package watch, no presence entities for Home Assistant, and no tracking. See External detection server.
If using a Coral or another accelerator directly is the point of your setup, Frigate’s documentation lists detectors for Coral, Hailo, Intel, AMD and Apple hardware. See Aperture vs Frigate and the comparison.
Limits
Section titled “Limits”- The timings are rough figures for a desktop CPU and a mid-range NVIDIA GPU, not guarantees for your machine.
- Aperture runs on x86-64 Linux only.
- Cameras must serve RTSP. Cloud-only cameras (Ring, Nest, Blink, most Arlo) have no local stream.
- The license is PolyForm Noncommercial: free for personal and other noncommercial use. See License.
Related
Section titled “Related”Object recognition · Requirements and sizing · Add cameras · Compare all NVRs