Warehouse Safety AI Analytics

The risk map is built from the incidents people bothered to report

Warehouse safety management depends on near-miss reporting, and near-miss reporting is unreliable everywhere. A driver who clips an upright, a picker who steps into an aisle as a truck passes, a reversing manoeuvre that ends closer than it should have — these are the events that precede serious accidents, and most of them are never written up. Not through concealment, usually, but because nothing happened, the shift is busy, and reporting takes fifteen minutes.

That leaves the site managing risk from a sample it did not choose. Racking damage is the clearest case: a strike on an upright reduces the capacity of a structure holding many tonnes above head height, damaged racking is a documented cause of catastrophic collapse, and it is precisely the event a driver does not report because the load stayed on the beam. Meanwhile pedestrian routes are drawn on the floor and crossed constantly at a few specific points nobody has identified, and an emergency exit gets blocked by a pallet every Thursday afternoon without anyone noticing the pattern.

• Most near misses are never reported, so the risk map is built from a biased sample.
• Racking strikes go unreported because nothing visibly failed at the time.
• Pedestrian and forklift conflict concentrates at points nobody has identified.
• Blocked emergency exits recur in patterns that manual inspection misses.
• Safety management is retrospective because the only data is incidents.

Solution overview

Swedish Technology applies computer vision to the specific events that precede warehouse accidents, and counts them. Pedestrian intrusion into vehicle areas, close-proximity events between people and trucks, excessive speed in aisles, and impacts against racking are all detectable at the right camera positions — and detecting them produces the leading indicator that near-miss reporting was always supposed to provide and never does.

Racking impact detection deserves particular attention because the consequence is disproportionate. A strike on an upright is a structural event, not a cosmetic one, and the usual sequence is that nobody reports it and the damage is found at the next scheduled inspection, if then. Automatic detection routes it to the racking inspection process immediately, with a clip and a location, so the assessment happens while the sequence of events is still known.

As with any monitoring of this kind, the design decision that matters most is what happens to the data. In a warehouse, a forklift is identified by its truck number, which means every event is attributable to a named driver by default — and a site that uses that for discipline will find drivers avoiding monitored aisles and reporting even less than before. We recommend intervention and location-based analysis, with individual attribution reserved for genuinely serious events under a policy agreed in advance.

How the solution works

1. 1
   Map where conflict actually happens Observe the site and identify the real intersections of people and vehicles — blind corners, dock approaches, cross-aisles — which is rarely where the floor markings suggest.
2. 2
   Place cameras for detection Positions and angles chosen for detection at those points rather than inherited from a security layout designed for coverage.
3. 3
   Detect the precursor events Pedestrian intrusion, close proximity, aisle speed, racking impact and blocked egress — each a documented precursor rather than a general safety metric.
4. 4
   Route by consequence A racking impact goes to the inspection process immediately with a clip; a recurring conflict point goes into the weekly review. Not everything needs an alarm.
5. 5
   Build the real risk map Events aggregated by location, hour, shift and activity, producing the picture that reported near misses never provided.
6. 6
   Act on the map, not the individuals Layout changes, barriers, mirrors, route redesign and supervision timing — interventions that address the location rather than the person who was there.

Key capabilities

Reference architecture

Edge processing, because a warehouse has many cameras, limited bandwidth, and no reason to send continuous video of its workforce anywhere.

Deployment options: Edge devices on site with a central dashboard. Detection continues through a network interruption with events buffered, since the immediate-alert path is the one with safety value.

Hardware options

Warehouse conditions favour vision analytics more than construction sites do — consistent lighting, fixed geometry, predictable traffic routes.

Swedish Technology supplies and integrates camera, sensor and edge equipment from established manufacturers, specified from a site assessment.

AI capabilities

Described by what it detects reliably and what it does not, since the difference is large.

• Person and vehicle detection — The foundation, and reliable in warehouse conditions. Distinguishing people from forklifts, pallet trucks and other equipment is well within current capability at typical camera positions.
• Proximity and near-miss classification — Measures distance and closing speed between people and vehicles to classify a genuine near miss rather than counting every co-presence. Threshold tuning per location is essential and takes a few weeks of real data.
• Racking impact detection — Detects contact events between vehicles or loads and racking structure. Low-speed contacts are harder to see than heavy strikes, which is why pairing vision with impact sensors on high-risk uprights improves coverage.
• Obstruction detection — Identifies pallets, equipment or stock blocking emergency exits and fire routes, including short-duration obstructions that inspections never catch.
• Speed estimation — Estimates vehicle speed in defined zones. Accuracy is sufficient to identify a pattern of excessive speed at a location and is not a calibrated measurement of an individual vehicle.

Integrations

These can be designed within project scope.

The integrations above are designed and implemented within project scope using vendor APIs, webhooks or standard connectors. They do not imply partnership, certification or endorsement by the system owner unless stated on that vendor's official pages.

Dashboards & analytics

• Risk map — Events by location, overlaid on the site layout — the picture that reported near misses were meant to provide.
• Racking impacts — Detected strikes with clip, location and inspection status, which should be reviewed as a structural matter rather than a driving one.
• Near-miss trends — Counts by location, hour, shift and activity, showing whether interventions changed anything.
• Egress compliance — Blocked exit events by location and recurrence, including the weekly patterns.

Security & deployment

Processing runs at the edge so video is analysed where it is captured and only events and short verification clips are retained. That keeps bandwidth realistic across a large camera count and avoids accumulating continuous video of a workforce. Detection continues through a network interruption with events buffered locally. Access to clips is restricted to HSE and management roles with a defined reason, and thresholds are tuned per location — a near-miss threshold that generates fifty events a shift in a busy cross-aisle will be ignored, and an ignored safety system is worse than none because it looks like coverage.

Data privacy

The attribution question is sharper in a warehouse than almost anywhere else in this solution set. Forklifts carry truck numbers and are assigned to drivers, so a detected event is attributable to a named person by default rather than by choice. That makes it a policy decision the site must make deliberately: our recommendation is location-based analysis for the overwhelming majority of events, with individual attribution reserved for genuinely serious matters under a policy agreed with the workforce in advance.

The practical argument reinforces the ethical one. A site that uses near-miss detection for individual discipline gets drivers avoiding monitored aisles and reporting even less than before, which destroys the leading indicator the system was bought to create. Workers should be told what is monitored and why, in languages they read, through induction and signage. Under UAE Federal Decree-Law No. 45 of 2021, employee monitoring needs a clear and proportionate purpose — and safety improvement at a location is one, while a general performance record is a different claim requiring a different justification.

Industry use cases

Implementation approach

1. 1
   Observation and conflict mapping Walk the site across shifts to find the real conflict points. Floor markings show the intended routes; observation shows the used ones, and the difference is the project.
2. 2
   Attribution policy Agree how event data will and will not be used, with the workforce, before deployment. Deciding this afterwards does not recover the trust or the data quality.
3. 3
   Camera assessment Establish which existing cameras have usable geometry and where new ones are needed. Warehouses reuse existing CCTV more often than construction sites do.
4. 4
   Baseline period Run in monitoring-only mode to establish the true event rate and tune thresholds per location. The initial count is usually far higher than the site expects.
5. 5
   Response design Define what is routed immediately, what triggers a racking inspection and what goes to weekly review. Not every event deserves an alarm.
6. 6
   Intervene and measure Change layout, barriers, mirrors or supervision at the highest-risk locations, then measure whether events fall — which is the point of the whole exercise.

Why Swedish Technology

• We treat near-miss counting as the product, because a risk map built from reported events is built from a biased sample.
• Racking impact detection is routed to structural inspection rather than to a driving conversation, because that is what the consequence warrants.
• We agree the attribution policy with you before deployment, since forklifts make every event individually attributable by default.
• Thresholds are tuned per location during a monitoring-only baseline, because an over-alerting safety system is ignored within a fortnight.
• The output is interventions at locations — barriers, sightlines, routes — not a list of drivers.

Limitations & prerequisites

• Detection accuracy depends on camera geometry. Well-placed cameras in warehouse lighting perform considerably better than outdoor equivalents, but poorly placed ones perform badly regardless of the model.
• Low-speed racking contacts are harder to detect than heavy strikes; pairing vision with impact sensors on high-risk uprights gives better coverage than either alone.
• Near-miss classification requires threshold tuning per location over several weeks. Before that, event counts are not comparable between areas.
• Speed estimation identifies patterns at a location; it is not a calibrated measurement of an individual vehicle and should not be presented as one.
• The system detects conditions, not causes. Whether events fall depends on the interventions made, which are layout and management decisions.
• Using event data for individual discipline degrades both cooperation and data quality. We recommend against it and design accordingly.

FAQ

Related solutions

• Interactive Experiences & Smart Visitor Solutions
• Warehouse Visitor & Contractor Management
• Loading Dock Management
• AI PPE Detection
• Truck Yard Management
• Smart Inventory Cycle Count
• Event Crowd Analytics System
• RFID Asset Management & Tracking