Detection is easy. Evidence is the product.
VisionX watches the cameras your plant already owns and hands back two things you can act on: a safety event with a thumbnail, a track ID and a timestamp — and a surface-quality decision with a disposition. It runs on your own hardware, and nothing leaves the machine.
Safety and surface quality, on the same feed
Every vendor in this category does one or the other. Running both through one platform means one deployment, one machine, and two lines of the P&L.
PPE, restricted zones, crowding, fire and man-down
Every person in frame gets a tracking ID, so one worker without a helmet is one event — not four hundred, one per frame. Zones are drawn live in normalised coordinates, so they survive a camera resolution change, and intrusion is judged on the worker's feet rather than the box centre — which is what makes it correct when someone leans over a barrier. Heavier layers like fire & smoke and fall detection toggle per source, so compute goes where it is worth spending.
One event for this worker, not one per frame — violations are debounced per track for 10 seconds. The thumbnail is the evidence, and it exports with the row.
Surface defects that escape a manual inspection
Hot and cold rolled surfaces are classified against a governed defect taxonomy, scored by defect area and class weight, and given a disposition. The confidence floor sits deliberately low: in surface inspection a missed defect costs more than a second look, and the faint hairline classes are exactly the ones a higher threshold drops. What the classifier cannot name, Anomaly Radar still flags.
Severity is a function of defect area and class weight, and every inspected surface leaves with one of four dispositions — plus a Pareto telling the mill which defect to fix first.
Connect, detect, decide, improve
No new cameras, no retraining before you can see anything, and no configuration step between plugging in a feed and watching it be analysed.
Connect
An existing CCTV feed, a recording, or a live IP camera over ONVIF — an address and a login, not an RTSP path.
Detect
Two detectors merged per frame, each person tracked, each call gated before it is allowed to become an accusation.
Decide
An event with evidence, or a surface with a disposition. Both export — CSV for events, Markdown for the shift report.
Improve
Uncertain detections queue for a human. Their corrections become the labels the next model version trains on.
For the questions a fixed class list cannot answer
A detector finds what it was trained to find. Four capabilities cover the rest — and each one is explicit about the limits of its own answer.
Ask the Camera
Type what you are looking for — “forklift”, “spill on the floor” — and search the current frame in plain words, including things no model here was ever trained on. It names which engine answered, so you know how far to trust the result.
Anomaly Radar
A trained model can only find defects someone already labelled. This scores how far a surface sits from the approved normal set and localises the strangeness on a heat map — the failure mode nobody has a class for yet.
Incident Reports
A shift-handover report written from the event log, not by a chatbot. Every count, cluster and corrective action traces back to a recorded detection with its evidence thumbnail. It exports as a document.
Review Queue
Everything the models were unsure about, waiting on a human. Confirm, correct or reject — and the next training round learns from this plant, in this light, on these cameras.
Absence of evidence is not evidence of absence
A false positive costs more trust than a late true positive — and in a plant it costs it in front of the people being monitored. So unknown is a first-class state everywhere, and the system says so out loud.
unknownA helmet-only model is never allowed to invent a vest violation. With no evidence either way the state stays unknown — it does not become an accusation.
calibratingAnomaly scoring refuses to return a number until it has enough approved normals to compare against. A confident guess would be worse than no answer.
which tier answeredEvery model has a documented lower tier, and the screen names the one that is currently answering. Nothing degrades silently.
The numbers, and where they came from
Benchmarked on a four-core laptop against a modest 4 GB mobile GPU — not a server. Detections are identical on both: the GPU buys speed, not accuracy.
| Stage | CPU only | With GPU | Gain |
|---|---|---|---|
| Safety detection (two models, merged) | 610 ms | 107 ms | 5.7x |
| Quality inference | 77 ms | 13 ms | 5.9x |
| Streamed video | 1.4 fps | ~9 fps | 6.4x |
Ninety-seven clips, every one verified by the detector
The bundled demo library is filtered by measurement, not by filename. After download, every clip is sampled with the real detector; any with no detectable person is quarantined. Eleven were — stock descriptions are unreliable, and “worker smoothing cement on a wall” turned out to be a macro shot of a hand. Eighty of the ninety-seven survivors contain genuine PPE violations.
Bring your own footage instead — drop in a recording and it is analysed with no configuration and no retraining.
- Construction27
- Manufacturing24
- Steel21
- Warehouse11
- Cement7
- Mining4
- General3
Nothing leaves the machine
Every comparable platform in this market is cloud or hybrid. VisionX makes no external call at runtime — the models, the database and the evidence thumbnails all sit inside your network, which is what makes it deployable on an air-gapped OT segment at all.
Connecting a live camera is an IP address and a login, resolved over ONVIF — no RTSP path to chase, no vendor lookup table, and no support ticket.
Bring us a recording from your plant
No configuration, no retraining, no data leaving your machine. We will run it in front of you and tell you honestly what the system can and cannot see on your cameras.
VisionX is a monitoring and alerting platform. It is not a safety-rated protective device and must not be used as a machine-stop interlock — that is a certified light-curtain or laser-scanner function under IEC 61496 and ISO 13855.