Crowd intelligence
Density, flow, compression, counter-flow and surge, sampled continuously. A density model counts crowds too dense to separate into individual people.
Crowd, behaviour, objects, identity and video events - all reading one shared picture of every frame.
Processing runs on your machine. Footage never has to leave the site; only conclusions travel.
Every incident carries the arithmetic behind its score. The system recommends, a human decides.
Bystander faces blurred before frames leave the machine; identity escalation needs approval.
Density, flow, compression, counter-flow and surge, sampled continuously. A density model counts crowds too dense to separate into individual people.
Bags are linked to the person who carried them. When one is left behind, the escalation carries the owner's last trajectory.
Loitering, counter-flow, running, erratic movement and restricted-zone entry - thresholded by your policy, not ours.
Authorised, audited watchlist matching across every camera in scope, answered in plain language.
Weakly-supervised models trained on public CCTV datasets, each reported with its measured confidence.
A live site map with camera coverage, and a field screen that sorts incidents by distance from the officer.
RTSP, USB or a recorded file, decoded on whichever GPU sits closest to the camera — the server here, or an operator's own machine acting as an edge node.
12 fps target per cameraYOLO11m locates people and objects; ByteTrack, written from scratch rather than pulled in, keeps an identity on each one as it moves.
7.5 ms detect on GPU, 54 ms on CPUAn OSNet appearance embedding per person, computed for the whole crowd in one batch, so somebody who leaves one camera and enters another stays the same person.
rank-1 92.8% on Market-1501Fifteen agents read the tracks: loitering, following, counter-flow, sudden running, crowd surge, an unattended bag, a face on the watchlist.
thresholds are yours, editable at runtimeAn incident carries the clip, the frames and the trail that produced it, and every read of it is written to the audit log.
nothing is asserted without its sourceAsk for one person and the gallery answers with every camera that saw them, in order, with the time they appeared at each.
a 15-minute window, your cameras onlyQuestions about what happened are answered from the incidents themselves, and every answer cites the one it came from. With no model configured it falls back to fixed wording rather than guessing.
answers carry their sourceBystanders are blurred before a frame leaves the machine, identity escalation needs an operator's approval, and footage past its retention window is deleted without being asked.
30-day default, yours to setMeasured, not claimed. 1080p with every agent running holds about 35 fps on a GPU and 11 fps on CPU. Those numbers come from benchmarking this pipeline, not from a datasheet.
It says when it does not know. Models are promoted only if they beat the baseline on held-out data — two were trained this month and refused. When a component is unavailable the platform reports the backend it actually used, so nothing implies a trained model where there is none.
Point it at the recorder you already have. Cameras, zones and thresholds are configured once and the agents run continuously from there.
8 cameras per node by defaultA machine with a graphics card can register as an edge node and take the cameras nearest to it. It polls outward over HTTPS, so nothing needs to be opened to the internet.
3x the throughput of CPU-onlyThe dashboard is a browser tab — no install, no plug-in. It draws the overlays locally, so a laptop can follow a site it is nowhere near.
the browser renders, the node infersBuilt for airports, railway stations, campuses and public venues — places with cameras already installed and nobody able to watch all of them at once.
Accounts are approved per organisation — an airport sees only its own cameras and incidents, never another site's.
CREATE AN ACCOUNT