01 · Detect
Object Detection
People, vehicles, pallets, tools, equipment, boxes and inventory — tuned per site.
Production-grade object detection, OCR, anomaly detection and safety intelligence. INNFINI Computer Vision transforms images and video into operational events that trigger workflows, alerts and evidence records — in real time.
Cameras and image streams flow through the vision pipeline — every detection becomes a structured event with evidence attached.
Video, image, mobile app, inspection tool, edge camera — the platform abstracts the source.
Identify people, vehicles, objects, equipment, documents, labels and safety conditions.
Plates, labels, serials, forms, shipment docs — extracted to structured fields.
Compare results against thresholds, workflows, expected quantities and system records.
Create alerts, incidents, inspections, exceptions or work orders — and stamp evidence.
Image, detection results, timestamp, model version — pinned to the audit ledger.
People, vehicles, pallets, tools, equipment, boxes and inventory — tuned per site.
License plates, labels, serials, forms and shipment documents — high accuracy under field conditions.
PPE violations, restricted-zone entry, unsafe behavior, overcrowding — alerted in real time.
Missing objects, wrong placement, congestion, damage or unusual activity — surfaced the instant the baseline shifts.
Attach images and detection results to incidents, inspections and audit records — automatically.
Run vision models locally where latency, bandwidth or privacy matters — sub-100ms on commodity gateways.
Count pallets, read labels, validate dispatch documents and detect mismatches at the gate before trucks leave the yard.
Detect vehicles, capture plates and link movement to yard workflows — fully hands-off.
Detect missing PPE and unsafe behavior in real time — alert supervisors with image evidence attached.
Traffic congestion, crowd density, parking violations and incident detection across the city camera network.
Counts, validations and exception reviews automated. Operators focus on the cases that need a human.
PPE adherence, restricted-zone enforcement and incident reduction measured at deployed sites.
Camera-to-RFID fusion on dock and yard operations — pixel-perfect counts.
Safety-critical decisions stay on-site — no cloud round trip required.
One figure for "vision accuracy" is meaningless — plate reading in daylight and PPE detection at night are different problems. These are measured per model family on deployed sites, at the operating threshold we ship.
| Task | Precision | Recall | Conditions | Failure mode |
|---|---|---|---|---|
| Vehicle detection | 0.97 | 0.96 | Daylight and lit night, to 40 m | Heavy occlusion in dense queues |
| Plate reading (LPR) | 0.94 | 0.89 | Under 60 km/h, plate within 25° | Damaged or non-standard plates |
| Person detection | 0.96 | 0.94 | Full and partial body, to 30 m | Crowds above ~4 people/m² |
| PPE compliance | 0.93 | 0.91 | Helmet, vest, eye protection | Back-turned subjects, dark PPE on dark ground |
| Pallet and carton count | 0.98 | 0.97 | Dock and gate, fixed camera | Stacked beyond three deep |
| Label and serial OCR | 0.96 | 0.92 | Print and laser-etched | Reflective film, condensation |
A generic model on your cameras will not reach the table above. Site tuning is the work, and it follows a fixed path.
| Stage | What happens | Typical duration |
|---|---|---|
| Camera survey | Angles, lighting, lens and occlusion assessed per camera. Some positions are rejected as unusable rather than tuned around. | 2–4 days |
| Baseline capture | Footage collected across shifts and weather so the model sees night, glare and rain — not only a good afternoon. | 1–2 weeks |
| Annotation and tuning | Site-specific classes labelled and the model fine-tuned. Your objects, your angles, your lighting. | 1–2 weeks |
| Shadow evaluation | Predictions scored against ground truth without triggering anything, until precision and recall hold. | 2 weeks |
| Threshold setting | Operating point chosen per task from the cost of each error — recall-led for safety, precision-led for enforcement. | 2–3 days |
| Go live and monitor | Continuous evaluation begins on day one. Drift raises an incident rather than degrading quietly. | Ongoing |
Vision on public or workplace cameras carries obligations that have nothing to do with model quality. These controls are defaults, not options.
| Control | Implementation |
|---|---|
| Process on-node | Frames can be reduced to events at the edge and discarded. Where this is enabled, imagery of people never leaves the site. |
| No face recognition by default | Person detection answers "a person is here", not "who". Identification is a separate, separately-authorised capability that most deployments never enable. |
| Redaction on export | Faces and plates are blurred in any image attached to an incident, unless a named role with a logged reason requests the original. |
| Retention by class | Raw frames, detections and events carry independent retention clocks — typically hours, months and years respectively. |
| Purpose binding | Each model is registered to a stated purpose. Using a safety model for productivity monitoring is a policy change with an approver, not a config tweak. |
| Full inference audit | Every detection records model version, threshold, confidence and source camera — so a disputed result can be reconstructed exactly. |
Computer Vision turns existing cameras into operational sensors — detecting, validating, alerting and pinning evidence into the system of record.
A 60-minute architecture review with our solutions team. We map your cameras, your workflows and the events vision should trigger.