Edge, IoT & Compute · Module
Vision AI
Existing cameras as privacy-preserving site signals that drive work
Vision AI is planned to run detection models on site edge nodes over existing RTSP/ONVIF cameras and send only metadata and blurred evidence upstream. Detections are reviewed and turned into EHS observations, store tasks and security alerts; face identification is off by default.
What it does
Vision AI turns existing cameras into site signals: footfall and queues for stores, PPE and restricted-zone presence for safety, shelf gaps for replenishment, smoke and flame indications as advisory alarms. It registers cameras from any RTSP or ONVIF source, runs models on the site edge node, and sends only detection metadata and blurred snapshots to the cloud. Frames are not persisted by default.
A privacy gate before any inference
No camera runs inference until its privacy profile is complete and approved: purpose, lawful basis reference, notice signage record, privacy masks, retention, allowed use cases and a DPIA link for high-risk use cases. Notice templates and data-principal request handling for footage are included. Evidence purge keeps running after cancellation; only a legal hold with a case reference pauses it.
A quality gate before any Action
Inference starts in review-required mode. Each site labels its own validation set, and precision and recall are reported per use case. Actions are enabled only when the Capability Brief threshold is met, and drift monitoring re-opens the gate. Reviewers confirm or reject detections with reasons.
From detection to work
Confirmed detections propose EHS observations, frontline tasks (restock, clean, open another counter) and security alerts through owner-app Actions, with the blurred snapshot, zone, time and model version attached. Detections describe site conditions, not identified individuals, and never drive person-directed disciplinary Actions.
What is never offered
Face identification, face templates, emotion recognition, gait or biometric identification, and worker-level detection timelines are not in this app. Any later face-identification evaluation (W6) would need a DPIA, counsel sign-off and an R4 admin Action, and is never offered for law enforcement or public surveillance. Clip retrieval needs privacy approval, is blurred by default, and unblurring needs a second approval with a reason.
Features
What Vision AI includes
Camera registry and discovery
ONVIF discovery and RTSP entry on the site LAN, credentials in the secret store, mapping to site, zone and location.
Camera privacy profile
Mandatory purpose, lawful basis, notice record, masks, retention, allowed use cases and DPIA link before inference.
Edge inference pipeline
Decode, mask, infer and emit metadata on the edge node; frames not persisted by default.
Zones and lines
Polygons, tripwires, dwell areas and queue areas per camera view, each with a purpose.
First-party use cases
People counting, PPE (helmet, high-visibility vest) and restricted-zone presence first; queue length, shelf gap and advisory smoke or flame next.
Detection events
Type, zone, time, confidence, model version, blurred snapshot, cooldown and deduplication.
Human review queue
Confirm or reject with reasons; required per use case until the precision gate passes.
Site validation gate
Per-site labelled validation set with precision and recall per use case; drift re-opens the gate.
Detection to Actions
Proposes EHS observations, frontline tasks and security alerts; never person-directed disciplinary Actions.
Evidence retention and purge
Short default retention for blurred snapshots, purge jobs and legal hold for incidents.
Camera health
Offline, obstructed, defocused, moved-view, low-light and tamper alerts.
Analytics dashboards
Footfall by hour and zone, conversion with POS aggregates, PPE compliance by zone, queue times, shelf-gap duration.
Bring your own model and model OTA
Import ONNX models with model card and licence under the same gate; canary rollouts through Box Manager.
Clip retrieval with approval
Short clips around confirmed incidents, blurred by default, with access logged and a second approval to unblur.
Workflow
Detection to operational action
- 1
Frames stay on the site LAN
- 2
Edge node masks, infers and blurs the snapshot
- 3
Metadata and blurred snapshot go to the Vision AI cell
- 4
Dedupe, cooldown and gate state are applied
- 5
Reviewer confirms or rejects when review is required
- 6
A task or EHS observation is proposed through owner-app Actions
- 7
Outcome and ROI entries recorded
Workflow
Clip retrieval
- 1
Investigator requests a clip around a confirmed incident
- 2
Privacy approver approves or declines
- 3
Blurred clip retrieved and stored restricted
- 4
Unblurring requires a second approval with a reason
- 5
Clip purged when retention ends
AI
AI skills and Digital Employees
Edge inference
PPE, restricted-zone presence, spill or obstruction, queue length, shelf gap and gate state on the edge node.
Privacy mask assistance
Proposes mask polygons for private doors, screens, name boards and street-facing regions.
Review queue ranking
Ranks by confidence, novelty and site risk and groups near-duplicates.
Camera health detection
Spots defocus, occlusion, repositioning, night-mode failure and tampering.
Metadata search
Natural-language search over detection metadata, never over raw video and never by face or identity.
Digital Employees that use this module
- AI Store Operations Assistant
- AI Site Safety Monitor (proposed)
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
Do we need new cameras?
No. Vision AI is designed to be camera-agnostic and work with existing RTSP and ONVIF cameras, with inference on a site edge node.
Does video leave our site?
By default only detection metadata and blurred snapshots go upstream. Frames are not persisted. Clips are retrieved only with privacy approval.
Can it identify people?
No. Face identification is off and not offered in W5. Detections describe site conditions and are never attached to a named person's HR record.
How do we know detections are accurate on our sites?
Each site labels a validation set; precision and recall are reported per use case, and Actions switch on only after the threshold is met. Drift re-opens the gate.
What about DPDP?
Purpose, notice, retention and data-principal requests are handled through Consent and Privacy Ops, with a DPIA for high-risk use cases. These are pointers; verify with counsel.
See Vision AI in your workflow
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.