The relevant industrial objects do not exist in standard datasets. That is why we build datasets from the real process and train custom detection models, with specific augmentation for day and night operation.
We have trained detectors for loaded and empty railcars; asphalt, aggregate, and fuel trucks; backhoe loaders; aircraft on the runway; and plant processes such as asphalt loading or boiler smoke. Every model is versioned and retrained when the environment changes.
On wood processing lines we classify logs by their defects — forks, protruding knots, curvature, over-diameter, abnormal ends — and detect foreign objects such as stones among the raw material. When the application requires dimensional measurement (diameters, distances), we do it with calibration against a known pattern and report the uncertainty: we never infer centimeters from pixels without validated calibration.
TRACKING, COUNTING, AND VIDEO ANALYTICS
On top of the detections we apply multi-object tracking to turn individual appearances into trajectories: we count railcars and trucks, determine the direction of travel, measure dwell times on the runway or in the plant, and group recurring visits.
- Counting by line or zone, with class discrimination (vehicle, person, machinery) and direction of travel.
- Dwell times per unit: how long each truck takes to load, each aircraft on the runway, each railcar on the siding.
- Trajectories and heat maps to understand how people, vehicles, and equipment move within the worksite.
- Recurring-visit grouping: the same unit detected N times is consolidated into a single visit with its history.
The result is reliable, traceable figures, with annotated video backing up every event.
LOG CLASSIFICATION BY DEFECT
On debarking and wood-preparation lines we operate a defect taxonomy defined together with the process owner, following the customer's quality instruction:
- Y/V-type forks and hammer ends (widened or protruding log ends).
- Protruding knots and localized bumps.
- Simple curvature and double curvature, measuring the characteristic distance of the inspection scheme.
- Over-diameter and under-diameter, against configurable thresholds (e.g. >45 cm).
- Stones and foreign objects on or between the logs: the class that causes the most damage to debarkers and chippers.
Each piece can accumulate multiple simultaneous defects: the event records all classes, their confidence, and their region of interest (box or segmentation). Thresholds are configured per class, line, and camera, and every change is audited — without recompiling the model.
RESPONSIBLE DIMENSIONAL MEASUREMENT
When the threshold is a physical magnitude —centimeters of diameter or curvature— an uncalibrated camera does not deliver measurement, it delivers an estimate. That is why we evaluate calibration against a known pattern, with proper optics and perspective control, and we always report three things: the measurement, its uncertainty, and the "not measurable" condition when the scene cannot guarantee it. We never infer centimeters from pixels without validated calibration.
EDGE OPERATION AND DEGRADED MODE
The classification engine runs on an edge server inside the plant, sized for 8–10 simultaneous cameras, and operates fully offline: inference and alerting do not depend on the Internet.
On camera failure, obstructed image, saturated GPU, or missing calibration, the system enters degraded mode and reports its unavailability — it never simulates an absence of anomalies. Every event keeps the original frame, the annotated frame, and a configurable clip, with local NTP time synchronization.
LATENCY AND OPERATOR ALARMS
An alarm's usefulness depends on the available reaction time: we compute the time budget as usable distance over transport speed, and measure end-to-end latency with capture, inference, publishing, and alarm-reception timestamps. The operator receives a visual and audible warning at the line; automatic process shutdown is left as a later stage, once the customer validates it.
Detection is only the raw material: the value lies in the rules that turn video events into decisions. On top of every detection we apply your operation's business logic and generate alerts, records, and reports with evidence.
IMPLEMENTED BUSINESS RULE TYPES
- Zone presence and absence: detecting that a truck arrived at the loading point, that a person is where they should not be, or that equipment is missing from its position.
- Maximum and minimum dwell time: alerting when a unit exceeds its expected time on the runway, in the plant, or at a process stage — or when the operation is suspiciously fast.
- Threshold counting: daily production, loading cycles, vehicles served; alerts when targets are exceeded or missed.
- Process sequences: validating that stages occur in the right order and within expected inter-stage times; any deviation is recorded.
- OCR reconciliation: cross-checking what the camera saw with the license plate or code read — the truck that loaded is the authorized one, the counted railcar is the declared one.
- PPE by zone and schedule: helmet, vest, and harness required per area; every violation generates a record with evidence.
- Time windows and calendars: rules active only on night shift, weekends, or maintenance windows.
- Rounds and patrols: verifying that the guard or supervisor completes their route, with pass time at each point.
NOTIFICATIONS
Each rule defines who to notify and through which channel: email with attached evidence, webhook/REST to your systems, and instant messaging (WhatsApp, Telegram) via gateway. The message includes the event photo or clip, camera, time, and the triggered rule.
ALERT ESCALATION
Alerts have severity and an automatic escalation flow:
- Every alert requires acknowledgment from the assigned responsible party.
- If there is no acknowledgment within N minutes, the alert moves up to the next level — supervisor, management — with staggered forwarding.
- Closure requires a comment; the entire cycle is audited.
- Response-time metrics per person and per event type, to drive continuous improvement.
FUTURE CAPABILITIES
- Composite rules: nested conditions combining multiple events and sources (video + telemetry + ERP).
- Self-adjusting thresholds: the system learns the operation's seasonality and adjusts its limits on its own.
- Deviation prediction: anticipating bottlenecks and non-compliance before they occur.
REPORTING
Everything measured is reported: live KPI dashboards, downloadable Excel spreadsheets, periodic PDF reports, and an API to feed your systems. This is the same reporting our customers use today in real operation.
Real data, real results
← Back to services Let's talk Request a POC
We offer scoped proofs of concept on your real operation, with agreed metrics and full integration upon completion.