Workplace safety and efficiency on the factory floor

Computer vision model trained by Neural to identify, on each worker, the presence or absence of Personal Protective Equipment — coveralls, gloves, safety glasses, masks and face shields — using the cameras already installed. Each non-compliance triggers an alert with photographic evidence and feeds safety metrics by area, shift and team.

How it works

1. 1The cameras (RTSP/IP) send images to the on-premises or cloud inference server.
2. 2The YOLO model detects people and each PPE item (or its absence) in milliseconds.
3. 3A tracker associates detections with each worker to avoid duplicate alerts.
4. 4Per-area rules define which PPE is mandatory (e.g. hard hat and vest on the construction site).
5. 5Alerts with photos are sent to the safety officer (dashboard, email, WhatsApp, webhook).

Benefits

• Continuous 24/7 monitoring, without relying on in-person rounds.
• Photographic evidence of every occurrence for training and audits (occupational safety regulations, ISO 45001).
• Compliance metrics by area, shift and team for targeted preventive action.
• Fewer accidents, lost-time injuries and labor liabilities.
• Leverages the existing camera (CCTV) infrastructure.

Productive × idle time statistics for workstations based on manual activity (hands and arms relative to the body). The system does not identify people by name: it measures the activity pattern at the workstation to support line balancing, team sizing and continuous improvement.

How it works

1. 1The pose model follows each worker's joints and the tracker keeps each person's identity.
2. 2Movement is measured at the hands relative to the person's own torso, with smoothed points and a dead zone. Body sway, shifting weight, walking, camera shake and small tics do not count as work.
3. 3Counted as productive: handling something for at least 2 s with hands away from the body and in front of/above the hips, or keeping an arm extended over a part/tool while making adjustments.
4. 4Hands hanging down, near the face (radio, phone) or resting on the body, as well as conversational gestures, count as a pause.
5. 5Short pauses between work cycles count as productive; with no manual activity for the configured time, the whole period becomes idle time.
6. 6Reports by workstation, shift and period reveal bottlenecks and opportunities.

Benefits

• Continuous time studies, with no timekeeper and no observer effect.
• Objective data for team and shift sizing.
• Identification of bottlenecks and waiting time for materials.

The AI locates the stack (or exhaust) and the smoke coming out of it, cuts out the exact outline of the plume and analyzes its color palette. Smoke color indicates combustion quality and what is being burned: white is usually just water vapor; dark gray and black indicate incomplete combustion (soot); bluish, burning oil; yellowish or brown, a sign of sulfur or NO₂. Darkness is measured on the Ringelmann scale, the same one used by environmental inspectors.

How it works

1. 1Text-prompted segmentation (CLIPSeg): in every frame the AI marks where there is smoke or vapor — without specific training.
2. 2Text-prompted detection (OWLv2) locates the stacks or exhausts and discards similar structures (distillation towers, plant buildings); the minimum confidence is adjustable.
3. 3The plume outline (not a box) is used in the analysis, so sky, buildings and the stack itself do not contaminate the colors.
4. 4Plume colors are grouped into the 4 dominant colors (k-means in CIELab space) and each is classified into a shade.
5. 5Darkness = how close the smoke is to black (0% = white, 100% = black). Ringelmann scale = darkness ÷ 20 (levels 0 to 5).
6. 6Each stack is analyzed separately (the smoke right above it; flare flames are excluded from color analysis). The indicators show the worst stack.
7. 7Persistent dark smoke (from the configured Ringelmann level) or abnormal coloring triggers an alert identifying the stack, with an image of the moment.

Benefits

• Smoke color reading → probable cause: vapor, particulates, incomplete combustion, oil, sulfur or NO₂.
• Automatic Ringelmann scale, the same used in inspections (e.g. black smoke limit for diesel).
• Continuous evidence for environmental permits and conditions (applicable emission standards).
• Immediate identification of incomplete combustion → fuel savings.

Automatic detection of water, oil, chemical and effluent leaks from cameras. The system identifies the source, measures the affected area and tracks how it evolves, triggering alerts for containment before the problem causes downtime, accidents or environmental damage.

How it works

1. 1Cameras cover critical areas (pump houses, tanks, pipes, docks, factory floors).
2. 2Text-prompted segmentation (CLIPSeg) marks the outline of jets, drips, runoff and puddles — without scene-specific training.
3. 3Regions that look more like vapor or smoke are discarded; the model trained by Neural reinforces detection.
4. 4The leak is confirmed in at least 3 of 5 consecutive frames, avoiding alarms caused by reflections.
5. 5The affected area is measured frame by frame to identify spreading.
6. 6Alerts with images are sent to maintenance and the occupational safety team.

Benefits

• Early detection reduces product loss and machine downtime.
• Prevents slips and falls — one of the leading causes of accidents.
• Avoids soil contamination and environmental fines.
• Historical records for predictive maintenance.

Virtual fences drawn on the camera image outline hazardous areas (machine working radius, presses, overhead cranes, docks, energized areas) or restricted areas (storerooms, perimeters). When a person enters, an immediate alert is triggered — and it can sound sirens or activate the machine interlock.

How it works

1. 1The user draws the zone on the camera image (free-form polygon).
2. 2The detector identifies people and vehicles; the position of the feet determines presence in the zone.
3. 3Each entry generates an event with photo, duration and time.
4. 4Webhook/PLC integrations make it possible to stop machines or trigger signaling.

Benefits

• Prevention of serious and fatal accidents involving machines and vehicles.
• Set up in minutes, with no construction work or physical sensors.
• Evidence for investigating incidents and near misses.

Human pose estimation (17 body keypoints) applied to every person in the field of view. When it identifies a sudden transition from standing to lying down — or a person down for a prolonged time — the system triggers an immediate alert to the response team, shortening the time to assistance.

How it works

1. 1The pose model locates each person's shoulders, hips, knees and ankles; the tracker follows each individual.
2. 2Standing: legs extended under the hips (knee above 140° and feet well below the hips) — bending over to pick something up still counts as standing.
3. 3Sitting: upright torso, thighs nearly horizontal and feet below the hips (chair, bench, edge of the bed). Leaning torso with legs extended = reclining (bed, armchair).
4. 4Crouching/kneeling: knees sharply bent with hips close to the heels, or body height much lower than the same person standing.
5. 5On the floor: horizontal torso without the legs supporting the body, or feet above the hips. Sitting on the floor: hips at the height of the feet.
6. 6Measurements are normalized by the person's own body size and posture is confirmed by voting over several frames, so it does not flicker.
7. 7Fall: a person seen standing who reaches the floor with a rapid drop of the hips. A person found already lying on the floor triggers an alert after a few seconds.
8. 8An alert with an image of the moment is sent within seconds; sitting, reclining and crouching do not trigger alarms.

Benefits

• Drastically shorter time between the fall and assistance ("long lie").
• Video records for root cause analysis and legal defense.
• Less visual-monitoring burden on the nursing staff.

Automatic analysis of images and videos from thermographic cameras for predictive maintenance and safety. The system converts the color palette into temperature, locates hot spots above the threshold and tracks the trend — without relying on a thermographer's manual interpretation at every inspection.

How it works

1. 1The color thermal image is converted back into intensity (inversion of the Ironbow/Rainbow/Grayscale palette).
2. 2Intensity is mapped to temperature based on the scale configured on the camera.
3. 3Regions above the alarm threshold are segmented, labeled and tracked.
4. 4With radiometric cameras, absolute temperature can be read directly through the manufacturer's SDK.

Benefits

• Predictive maintenance — failures identified weeks before breakdown.
• Prevention of electrical fires.
• Continuous inspections instead of monthly rounds.

Detection of the presence and use of cell phones in environments where they are prohibited for safety, quality or confidentiality reasons — such as production lines, machine operation, cleanrooms, exams and areas with sensitive information.

How it works

1. 1The detector identifies people and cell phones in each frame.
2. 2The phone is associated with the nearest person (hands/face).
3. 3The occurrence is recorded per tracked person, with accumulated usage time.

Benefits

• Reduces distraction-related accidents.
• Enforces information security policies.
• Objective records for feedback and training.

Facial monitoring of drivers and operators to identify signs of fatigue and microsleep — eyes closed for a prolonged time, rising PERCLOS, frequent yawning — sounding alarms and notifying the control room before an accident happens.

How it works

1. 1The face mesh (468 points) is computed on every frame.
2. 2The EAR (Eye Aspect Ratio) measures eye openness with a per-person adaptive threshold.
3. 3PERCLOS indicates the percentage of time with eyes closed (NHTSA standard for drowsiness).
4. 4MAR detects yawns; the combination of signals determines alert, fatigued or drowsy.

Benefits

• Prevents accidents caused by drowsiness — up to 20% of serious road accidents.
• Immediate alerts at the workstation and in the monitoring center.
• Fatigue metrics by shift for working-hours management.

Automatic verification of jewelry use — rings, wedding bands, watches, bracelets, necklaces and earrings — in environments where they are prohibited due to biological risk, product contamination or mechanical safety. It uses an open-vocabulary detector: new classes can be added by text, without retraining.

How it works

1. 1The YOLO-World detector receives the list of jewelry items in natural language.
2. 2Each item is associated with the person wearing it.
3. 3Non-compliance triggers an alert with the image crop as evidence.

Benefits

• Auditable compliance with occupational safety regulations and GMP.
• Less cross-contamination and fewer product recalls.
• Prevention of entanglement and degloving injuries.