A LiDAR + camera pair detects when a worker enters a hazardous area — ATEX zone, crane swing radius, conveyor pull-cord zone, automated cell — and feeds slow-down or hard-stop signals into the safety PLC before the worker is at risk. Works in dust, fog and low light. Every event is logged for the safety officer's audit trail. Custom-built for your zone topology and your plant safety logic.
ATEX worker safety AI is a perception pipeline that detects people entering hazardous zones — ATEX areas, crane swing radius, conveyor pull-cord zones, automated-cell perimeters — and triggers PLC slow-down or hard-stop signals before the worker is at risk, with an auditable event log for the safety officer.
The pain is structural in every dust- or gas-prone plant: existing ATEX safety relies on light-curtains, RFID badges and procedural lockout-tagout. Light curtains miss anyone who walks around them; RFID badges only protect the worker who actually wore one; lockout-tagout is bypassed under schedule pressure more often than safety departments care to admit. And when an incident does happen — a near-miss, a swing-radius excursion, a pull-cord trip — there is rarely a clean record of who was where, when, and why the protection failed. Dust, smoke and low light defeat camera-only systems exactly when you need them most.
Our approach fuses two sensors per zone. A 3D-LiDAR scans the zone geometry — point clouds penetrate dust and smoke far better than RGB, so the system still sees a person when the camera does not. A high-resolution camera adds confirmation — rules out forklifts, animals and pallets that would otherwise trigger false stops. Per-zone safety distance is configured per area (a crane swing radius needs a larger margin than a robot-cell perimeter), and the redundant detection writes interlock signals into the safety PLC. Every event — detection, slow-down, stop, all-clear — goes into an audit log with timestamp, zone, person count, sensor confidence and evidence frames.
This is not an off-the-shelf safety appliance. ATEX zones, crane envelopes and plant safety logic differ per site. We treat every engagement as a discovery + fixed-scope build, descending from our hub service Industrial Perception AI. Typically four to twelve weeks from contract to live tool, with the code and the model handed over to you at the end.
Built by the same team that delivers our LiDAR wagon-detection pipeline — six lid centroids, sub-5 cm error, two-second response.
One or more LiDAR + camera pairs cover the hazardous zone with redundant overlap, so no single sensor failure leaves a blind spot. ATEX-aware sensor housings where the zone classification requires it (dust zone 22, gas zone 2, fertiliser silos, refinery process areas). Dust-tolerant optics, runs in fog, smoke and night light — the geometry channel keeps working when the visual channel struggles.
LiDAR clusters persons by geometry — height, gait, motion signature — so the system separates a worker from a forklift or a pile of pallets. The camera confirms the detection visually, ruling out animals, vegetation and yard equipment that would otherwise trigger false stops. Per-zone safety distance is configured: crane envelopes need larger margins than a robot-cell perimeter.
Confirmed detections feed into the plant safety PLC as slow-down or hard-stop signals over the existing safety bus. Every event — detection, slow-down, stop, all-clear — writes into the historian or safety database with timestamp, zone, person count, sensor confidence and evidence frames. The safety officer gets an auditable trail without manual paperwork.
We keep the architecture boring on purpose. Three loosely coupled stages, each one independently testable, each one swappable when the sensor stack changes. Built on the same stack we use across all our perception work: PCL, Open3D, OpenCV, PyTorch.
Each LiDAR delivers a 3D point cloud of the zone — typically several hundred thousand points per scan, multiple times per second. Cameras add synchronised colour frames, lens-calibrated and time-stamped. Raw data writes to local storage with a per-event UUID for audit trail and re-labelling. The LiDAR channel is the resilience layer: it keeps reporting geometry when dust, fog or low light blind the camera.
The point cloud goes through ground removal, voxel down-sampling and a person classifier trained on real plant data — workers in PPE, helmets, high-vis vests, not the office benchmark. The camera image goes through a second classifier on the same target. Both heads fuse into a confirmed person track with a measured distance to the configured zone boundary. False-positive rate is tuned per zone; the safety distance is configurable per area.
Confirmed track crossing a zone boundary triggers a slow-down or hard-stop signal into the safety PLC over the existing safety bus — Profisafe, CIP Safety, or hard-wired safety outputs per project. Simultaneously the event writes into the historian or safety database: timestamp, zone, person count, confidence, evidence frames. The control-room operator sees the live overlay on the HUD; the safety officer queries the database for the audit trail.
All three stages run on an industrial PC at the zone cabinet. No cloud dependency, no external API, no licence dial-home. The code is yours at handover.
Slow-down and hard-stop signals into the plant safety PLC, over the safety bus or hard-wired safety outputs. The safety logic stays under your existing functional-safety regime — we feed the inputs; your safety PLC owns the decision and the certification scope.
Every detection logged with timestamp, person count, zone, sensor confidence and evidence frames. Feeds your safety database, your historian, your incident-investigation workflow — the same temporal model we use on stockpile monitoring, applied to safety events instead of material piles.
A live overlay so the control-room operator sees who is where in real time — zone outlines, active detections, current safety state per area. Same dashboard surfaces the recent event history for shift handover and supervisor reviews.
Every plant has its own ATEX zone topology, its own crane envelopes, its own safety PLC and its own lockout-tagout culture. A generic safety appliance solves the generic case; your case is not the generic case.
A crane swing radius needs a larger margin than a robot-cell perimeter; a fertiliser silo with explosive-dust risk needs a different reaction profile than a wood-pellet conveyor. We configure the safety distance, the reaction curve and the false-positive tuning per zone — not one global threshold for everything.
ATEX-aware sensor housings where the zone classification requires it — dust zone 22 in grain and pellet handling, gas zone 2 in refinery and chemical, dust-and-gas hybrids in fertiliser, IP66+ wash-down in foundry. We pick the housing per zone, not one catalogue option for the whole plant.
Most plants already have a safety PLC, light curtains, RFID badges and a lockout-tagout procedure. We feed our detections into what's already there as additional inputs — we do not replace your functional-safety regime. Profisafe, CIP Safety, hard-wired safety outputs and OPC UA for the non-safety audit channel all supported.
You own the source code, the model weights, the labelled dataset and the audit log at handover. The audit log lives in your historian, not on a vendor cloud. We document the system, train your team, and walk away clean. No black box, no monthly per-zone licence, no service contract you can't exit. See our FAQs on IP and engagement model for the standard terms.
The engagement-model questions we hear from every plant safety officer considering a custom perception build. Need something more specific to your zone? Ask us.
Zone classification, sensor setup, your safety PLC interface, your existing lockout-tagout procedure — anything you have. We come back within two business days with an honest first assessment and a fixed-price scope for the Discovery workshop.