One process per camera
Each video channel runs as its own Windows process. No Windows sign-in is required to keep it running.
On-premises camera analytics
CrowdSense Vision Engine
Windows software that sits next to your cameras. It reads people and vehicles on the machine, then sends only the events (line crossed, zone joined) off site. This is the spec we walk in demos.
Service 3.x · one 1080p stream equals one channel
Metadata only on the wire
The pictures stay on the machine. Only events (line crossed, zone joined, and the like) leave the site, so a thin or unstable link still works.
Tune it on the box
A live view and a recent-events list sit next to the stream, so lines and zones can be set and checked on site without waiting on the cloud.
Detection profiles
Pick the profile for the job. AI is more accurate and classifies what it sees. Non-AI is cheaper on the CPU and takes more cameras.
| Detects | Accuracy | Classification | Speed | |
|---|---|---|---|---|
| AI profiles | People, vehicles or faces | Up to 98% | Person, bike, car, van, truck, bus. Face age and gender on the faces profile. | 20 ms per frame |
| Non-AI profiles | Any moving object | Overhead up to 92%. Tilted up to 80%. | None | 2 ms per frame |
- AI profiles. Machine-learning models. Fewer false detections from shadows and sun. Needs more CPU, and AVX2.
- Non-AI profiles. Background extraction. Lighter on the machine. More channels per box.
Overhead
0–20°
Looking down. Used with the overhead property on some profiles.
Tilted
20–70°
Looking along the scene. Used with the tilted property on some profiles.
Hardware, 1080p channels
One channel is one camera stream. Figures are a ceiling for a typical 1080p feed, not a promise for every scene.
| Machine | AI profiles | Non-AI profiles |
|---|---|---|
| Intel i9, 64 GB RAM, GeForce RTX | 30+ channels | 40–80 channels |
| Intel i7, 32 GB RAM | 7–15 channels | 25–35 channels |
| Intel i5, 16 GB RAM | 5–10 channels | 17–25 channels |
| Intel i3, 8 GB RAM | 3–6 channels | 10–20 channels |
| Intel Celeron / Atom, 4 GB RAM | Not supported | 3–7 channels |
Worked example. Intel i9-14900K (3.20 GHz), 64 GB RAM, Nvidia GeForce RTX 4070. Tested at 30 channels.
- Windows 7, 8, 10, 11 and Server.
- Non-AI profiles: 32-bit or 64-bit.
- AI profiles: 64-bit only, and the CPU must support AVX2.
- AMD CPUs follow the same memory and channel bands as the Intel table.
- A discrete GPU is not required. It is recommended when you want the top of the channel range, because it decodes video faster.
Cameras
What it will talk to
Any webcam. Any IP camera that can stream RTSP. ONVIF is not required. It is useful for discovery.
Resolution
Any resolution works. Smaller frames (around 360p) are easier on the box, as long as neither side is under 256 pixels. For the faces profile, use 720p or 1080p.
Tilt
Overhead profile: 0–20°. Tilted profile: 20–70°. AI people and vehicles: 20–90°. AI people: 0–90°. AI faces: 80–90°.
Smallest object the engine will reliably see
| Case | Minimum width in the frame |
|---|---|
| AI, vehicle front or back | 4% of frame width |
| AI, vehicle side-on | 10% of frame width |
| AI, person | 3% of frame width |
| AI, face | 4% of frame width |
| Non-AI, any moving object | 3% of frame width, absolute minimum |
Example: on a 1280×720 frame a vehicle seen from the front needs to be at least 50 pixels wide.
Bring this to the next site walk
If the cameras, the box and the angles fit this page, we can talk about a first event. If they do not, we will say so before anyone unboxes a node.