A forklift emerges from behind a rack just as a pedestrian steps around a pallet. A plant planning to install AI cameras for forklift–pedestrian proximity detection must solve that specific encounter: reduce the crossing first, observe what remains early enough, warn the right people, and define the response. A good person-detection score alone cannot reveal someone fully hidden by a load, and an alert that the driver cannot hear changes no behavior. This guide focuses on commissioning warnings at blind corners for driver-operated forklifts in factories and warehouses, from the site survey to acceptance testing.
Start with one crossing and one defined response
Choose a single crossing where a forklift route and a pedestrian route overlap. Before proposing cameras, ask whether routes, crossing points, barriers, speed limits, mirrors and stopping rules can prevent the encounter. Then use an AI camera to supplement the controls where a residual blind corner remains. OSHA’s forklift guidance calls for separating forklift and pedestrian traffic where possible, slowing or stopping at obstructed intersections, and using a spotter when visibility requires one. The UK HSE also prioritizes separate routes and visible crossing points. These are design references, not statements of Thai legal requirements. OSHA pedestrian traffic · HSE route separation
An ordinary vision system observes, warns and records. It is not automatically a safety-rated protective function. If an automatic reduction of speed or stop is contemplated, competent specialists must design and validate the entire vehicle and control-system safety function, including communication failures, braking distance and restart. A camera demo or an average accuracy number does not prove that function.
| Decision | Pilot example | What goes wrong if it is undefined |
|---|---|---|
| Hazard scenario | A person and a driven forklift approach a rack-obstructed T-junction | Normal passage is confused with hazardous approach |
| Primary control | Separate walkways, designated crossing, speed and stop rules | AI becomes a substitute for traffic design |
| Observation | Both approaches visible before the conflict point | The warning arrives too late |
| Alert recipient | Driver-side and pedestrian-side indicators, with supervisor escalation where needed | Detection produces no action |
| Response | Driver slows and stops safely; pedestrian waits at the crossing | Alarm volume rises without behavioral change |
| Fault state | Loss of observation is shown and a fallback traffic rule begins | Silence is mistaken for safety |
Our factory safety AI camera RFP and acceptance guide covers wider procurement, governance and privacy questions. This article narrows the work to forklift–pedestrian warning at one blind crossing. It does not address PPE monitoring or visual quality inspection. Our separate automated forklift guide concerns driverless vehicles; its protective-function design should not be copied onto an advisory system for driven forklifts.
Map the conflict before selecting a camera
On a site plan, mark normal and reverse vehicle movements, loading positions, pedestrian shortcuts, doors, pillars, racks and temporary pallet storage. Observe shift changes, cleaning, replenishment and visiting drivers rather than only the quiet demonstration hour. Ask drivers, pedestrians, maintenance and EHS staff independently where they lose sight, where horns disappear into noise and which loads block a sight line. Use crossing frequency, speed, available visibility, obstructions, stopping space and near-miss reports to prioritize the first location. NIOSH’s forklift alert reviews fatal events and prevention measures; its cases help prompt local questions, but they cannot be converted directly into a probability for your plant. NIOSH forklift alert
Record shift, truck type, load height, direction, pedestrian entry point, lighting and actual crossing behavior. Collect only the information needed for traffic design. Before saving sample footage, agree on notices, access, retention and purpose with the site’s privacy and labor stakeholders. A safety camera should not quietly become a system for scoring an individual’s work rate.
First search for ways to remove the crossing: move storage, use a different door, or separate delivery and shift-change periods. Second, examine barriers, a protected walkway, a designated crossing, mirrors, a speed rule and a stop line. Third, ask whether the remaining encounter can be seen soon enough for a warning. The most concerning location may be unsuitable for a camera pilot if a tall load hides a pedestrian until the last moment. In that case, change the route or layout. OSHA’s hierarchy of controls favors eliminating or engineering out a hazard before relying on procedures or PPE. OSHA hazard prevention
Place the camera where an approach becomes visible early
A camera directly above the middle of a crossing may see both parties only when they are already together. At each proposed mounting point, test the approach view, rack and pillar occlusion, load height, raised forks, other trucks and lighting transitions. Mounting higher is not always better: people may occupy too few pixels and disappear behind tall loads. Mounting lower can put pallets across the entire view. Assess how early and how continuously each party is visible, not merely whether the intersection appears in a still image.
Do not set a universal warning radius in meters. Work backward from the stopping opportunity at this location. Conceptually, the needed lead distance is vehicle speed multiplied by detection, transmission and human response time, plus a braking distance measured for the truck and conditions, plus a margin. Truck type, load, floor, slope and driving practice alter those terms. Measure in safely controlled conditions and incorporate the results into the site’s risk assessment. OSHA’s travel guidance calls for speeds that allow safe stopping under all travel conditions. OSHA traveling and maneuvering
If one camera cannot see both approaches, compare two views, a different crossing location and stronger physical traffic controls. Two cameras add time synchronization, duplicate-object handling and detection of either link failing. Overlapping fields of view may still be blocked by the same load. Mark each camera’s blind area on the plan and update it with representative load sizes.
Test the actual video stream across shifts: low light, backlighting at a doorway, floor reflections, rain near entrances, reflective vests, vibration, dirty lenses and changes in door position. A high-resolution recording setting is no assurance that a compressed, delayed stream contains enough detail for moving-object detection. Check field of view, frame delivery, end-to-end latency and access to the usable stream. If a person is absent from the image, retraining the model cannot make that person visible. Add light, change the camera or change the route.
Separate occlusion into fixed (walls and rack ends), variable (pallets, parked trucks, loads and doors), and mutual (one person or vehicle hiding another). Fixed occlusion suggests relocation or route redesign; variable occlusion requires storage discipline or another view; mutual occlusion may require an explicit uncertain state and conservative operating rule. Put all three in the acceptance test. A partially visible person is not automatically detectable.
Turn object detection into a meaningful proximity event
Person detection alone will alert on a worker walking safely inside a protected lane. Forklift detection alone may alert every time a vehicle parks or loads. A useful event considers people, vehicles, direction of travel, permitted routes, relative position, duration and visibility. A single camera cannot necessarily provide reliable real-world distance or speed. Perspective correction requires floor reference points and stable mounting; load height and orientation can still change error. If a displayed distance cannot be validated, define a zone-entry rule rather than pretending to measure precise separation.
One candidate rule is: a pedestrian is in the designated crossing zone while an approaching forklift enters its approach zone and the condition persists long enough to warrant a warning. Tune persistence with local tests of missed and repeated warnings. Classify parked vehicles, safe parallel movement and normal walkway use so alarms remain useful. Do not suppress so aggressively that a person emerging from behind a pallet, or a second person arriving during a suppression interval, is missed. Model confidence is only one input; consider the consequence of a miss. NIST’s AI Risk Management Framework stresses evaluation of validity, reliability and safety in the context of deployment and across the system life cycle. It does not prescribe a forklift-camera accuracy target. NIST AI risk and trustworthiness
Design the alert around human action
A red box on a screen does little if the driver is looking into the turn. A louder siren can disappear among reversing alarms, machines and hearing protection. Decide separately what the driver and pedestrian need to know, where they will be, and how they will recognize the warning. Test candidate vehicle-side lights, pedestrian-side indicators, audible signals and supervisor escalation under actual noise and illumination. Do not rely on color alone. Train visitors as well as regular staff in the signal’s meaning.
Write the response in plain operational terms. For example, the driver recognizes the warning, reduces speed, stops where stopping is safe, and verifies the pedestrian’s position. The pedestrian waits at the designated crossing and checks that the vehicle has stopped. State who can clear the event and resume traffic. The absence of an alert is not permission to proceed. Existing look, horn, stop and spotter practices continue. OSHA recommends slowing, stopping and sounding the horn at obstructed intersections; an added camera must not weaken those practices. OSHA pedestrian traffic
Measure false alerts by their effect, not only by a daily count. Ten alerts in a minute can encourage people to ignore them more than ten alerts across a shift. Group duplicates only under a documented rule, and ensure a new pedestrian still triggers a warning during any cooldown. Record total alerts, verified hazardous approaches, whether recipients noticed the alert, and interruption time. Classify false alerts as object error, spatial error, overbroad zone, parked-vehicle alert, duplicate event or a mismatch with the site rule. Classify misses as occlusion, darkness, out-of-view entry, dropped frames, latency or model error. Each calls for a different correction.
Treat an unobservable scene as a fault, not an all-clear
Dirty or moved lenses, a pallet in front of the camera, frozen video, a disconnected network and a drifting clock do not mean that pedestrians have vanished. Monitor frame freshness, connection, time, last good image and field-of-view changes. Distinguish no hazardous event observed from the scene cannot be observed. Send a fault to the responsible site role and start the fallback traffic rule selected by the risk assessment. Depending on the crossing, that may mean manual control, stricter access, reduced traffic or a temporary pause. Never show a green all-clear simply because the camera is offline.
After a link recovers, separate historical events from current warnings so old alarms are not emitted as a fresh burst. Test storage full, maintenance bypass and configuration restoration after an update. Make rack, uniform, vehicle and lighting changes triggers for reassessment. Provide a clear route for a worker to report a suspected missed person and, when needed, activate fallback controls pending investigation.
If automatic braking is requested, scope it as a separate engineered safety function. Define the required risk reduction, vehicle interface, communication loss, delay, erroneous stops, stopping distance and restart. A qualified team must determine applicable requirements and validation. An abrupt stop can itself create a load hazard. Neither a successful vision demo nor this article’s advisory warning criteria prove a complete braking function.
Factory acceptance and site acceptance: test the encounter
Use FAT to check system configuration, event rules, alert routing, logs, permissions and injected communication faults. Use SAT to test the actual occlusion, lighting, noise, routes and people at the site. Do not create a dangerous near collision to prove the device works. Conduct controlled tests at safe speeds and separations, under a documented plan with trained participants and restricted access. Use safe mock walking, replay or signal injection where necessary, but label simulated results separately from observations with a real vehicle.
Each test record should connect start condition → input → expected detection → recipient → human response → evidence and decision. Agree pass limits based on available stopping opportunity and a tolerable operational alert load, not a vendor brochure. A bare statement such as “alert within X seconds” is meaningless if a pallet prevents observation until contact. Record first visible location, truck speed, load, lighting and delivery time together.
| SAT scenario | What to exercise | Evidence to retain |
|---|---|---|
| Normal approach | Both parties move toward the named crossing | Detection, delivery, recognition and stop times |
| Fixed occlusion | A person emerges from a rack or pillar | First visible point and warning point |
| Variable occlusion | Representative load and a parked pallet | Blind area and fallback rule |
| Reverse or alternate route | Permitted movements other than standard travel | Correct direction rule |
| Harmless passage | Person inside walkway, parked truck, safe parallel movement | Unnecessary warnings and their burden |
| Lighting | Each shift, backlight and doorway contrast | Video and rule outcome |
| Multiple actors | Overlapping people or trucks; another person during cooldown | No missed independent event |
| Observation fault | Obstructed lens, frozen frames, disconnected link | Fault signal, fallback and recovery |
| Change | Camera move, new rack or model update | Regression test and approval |
Do not buy a claimed “99% accuracy” without its denominator. Track hazardous-scenario detection, false-alert rate, delivery time, fault detection and actual human recognition separately. Define an event, the number of trials, which shifts were covered and what happens when settings change during a test. Frame-level person classification is not the same as successful warning for a dangerous crossing. A small sample with no misses cannot establish that rare scenarios never fail. Make important occlusion and fault cases mandatory, then observe warning load during a defined operating trial. Re-run the relevant tests after change and periodically in service, consistent with NIST’s emphasis on ongoing evaluation. NIST AI risk and trustworthiness
Write the procurement request around evidence
Give bidders a site drawing, current traffic rules, barriers, truck and load types, directions, lighting, network conditions and data policy. Request a proposed camera position with marked blind areas, lighting assumptions, event rules, alert destination, fault indication, SAT plan, and responsibilities for maintenance and changes. “Can detect forklifts” is not a purchase specification. State which scenario should produce which signal to whom, early enough for which action.
Separate the costs of cameras, mounts, cables, lights, edge computers and indicators from survey, traffic-route works, risk assessment, rule tuning, SAT, training, maintenance, fallback procedures, retesting and video governance. Check latency and fault visibility even if the existing network appears adequate. Budget lens cleaning, spare parts, recalibration after a move and regression tests after model updates. Agree access and retention before footage is collected. If individual identification is unnecessary, favor aggregate event records and limited retention.
Ask vendors to draw where a person first becomes visible under your load conditions, rather than promise a detection distance from a generic brochure. Ask which hazards will be replayed in each shift and which miss causes failure, rather than accept one average score. Ask whether drivers recognized the warning in plant noise and what happens if they did not. Ask how a blocked lens differs from an uneventful crossing and who is told to apply a fallback rule.
Pilot sequence and a conservative business case
Begin with route observation and removal of avoidable crossings. Pilot one residual blind corner with both hazardous and harmless examples. Then run SAT, confirm training and assign an owner for each fault and response. Expanding to another crossing requires a new view and route assessment; copied settings are insufficient when racks, doors, loads and lighting differ.
Avoid claiming a return on investment from an invented accident-reduction percentage. Severe events are infrequent, and a short, single-site pilot cannot estimate their reduction reliably. List total life-cycle cost alongside alert review time, avoidable interruptions, traffic-rule adherence, near-miss reporting quality and system availability. Convert only measurable items to money; retain safety significance as a separate decision factor. For example, alert count multiplied by average review time and operating days can estimate labor demand, but minimizing alerts to zero is not the goal. Compare at least three options: route changes alone, physical separation with existing rules, and physical separation with AI warning. The HSE’s route-separation guidance explains why layout belongs in the comparison before hardware. HSE route separation
Once live, a falling alert count might reflect fewer encounters, a changed route, stricter software thresholds, or a camera that no longer sees. Review vehicle and pedestrian traffic, observable time, rule changes and blind areas alongside the count. Daily checks cover lens, mounting, view, clock, indicator, link and storage. Periodic review covers near misses, worker feedback, false and missed alerts, and whether an engineering change can remove the conflict altogether. New racks, pallet positions, clothing, lights, trucks and software versions should trigger the relevant regression tests. If drivers stop responding, examine alert timeliness, frequency, audibility and work pressure as well as training.
What the first site walk should produce
Do not end the first walk after finding a convenient column for a camera. Produce a floor plan with vehicle and pedestrian approaches in different colors and watch the forklift both loaded and empty. If its rear swings toward a walkway during a turn, record that encounter as well as the front approach. Separate nominal aisle dimensions from current conditions: rack overhang, temporary pallets and worn floor markings can change the usable route. Check with facility owners where a crossing could move and how any change interacts with emergency exits and fire equipment.
Ask workers to describe in ordinary words who approaches from which direction and what makes the encounter hazardous. Disagreement between drivers and pedestrians can reveal an ambiguous traffic rule. A driver may intend to stop before a corner while a pedestrian has no reason to expect that stop. The camera project should expose and resolve this gap, not conceal it with an alarm. Collect footage across representative loads and shifts. Record the capture date, view, lighting and settings. Separate footage used to adjust a model from footage used to test it; otherwise a system can look unusually strong on images it has already encountered.
Stopping opportunity is more than a distance on a drawing
The lead-distance formula aligns a procurement discussion; it does not create an acceptance threshold by itself. Measure actual approach speed, loaded visibility, floor condition and where a driver first notices the signal, not only a truck’s nominal top speed. A light beyond the driver’s line of sight delays action even if the zone is far away. A pedestrian signal may help where vehicle noise is masked, but a startling signal that makes someone step into the aisle would create a new hazard. During the test, verify that the person can remain in a safe position after the warning.
Likewise, a stop point may be unsuitable if it is on a slope or directly in front of a shutter, emergency exit or intersecting vehicle route. Warning zone and stopping behavior form one traffic design. A device that first alerts after a truck enters an area with no safe stopping opportunity may have excellent object detection but fail as a warning. An overly early alarm that sounds constantly becomes background noise. Compare candidate mounting positions and event zones under the same driving conditions, recording alert time and actual response. Preserve rejected options and their blind areas so a later layout change can be reviewed intelligently.
Make ownership legible across vendors and shifts
A software supplier may own the model, an electrical contractor the wiring, maintenance the camera and EHS the traffic rule. That split creates gaps during a fault. Make a responsibility table for lens cleaning, field-of-view checks, clock synchronization, indicator testing, threshold changes, log review, worker notice and retesting. Name both performer and approver. Record who receives a fault at night or on a weekend, who substitutes when that person is away, and who authorizes normal traffic after repair. A support contract should state the fallback traffic rule and evidence of recovery, not merely a help-desk response time.
For a model or rule update, retain the old settings and connect the reason, effective date, affected cameras, regression results and approval. A change that reduces nuisance alarms must not be released if an important occlusion test now fails. If staff can temporarily mute an alarm for maintenance, log who did so, why, the fallback measure and its end time. “Under maintenance” alone does not tell workers whether to rely on the system. Show normal observation, observation unavailable and alerts disabled as distinct states.
Hand over a usable operating package
The deliverable is more than a camera list. Hand over the current crossing plan, each field of view and blind zone, included and excluded scenarios, approved traffic rules, alert meanings, fault fallback, FAT and SAT results, configuration and model versions, and a maintenance checklist. Specify who may view safety footage and for how long. In training, show not only a successful warning clip but also a person hidden from view, harmless passage that should not alert, and the display for a system fault. Teaching only what the system can do invites overconfidence when it cannot observe.
At the first post-launch review, look beyond alert counts. Has route use changed? Are stop points observed? Is any signal difficult to notice? Ask drivers and pedestrians which display is unclear and when avoidable waiting occurs. Before changing detection thresholds, examine whether a crossing or pallet-storage rule should change. Keep the reason for each decision. The next crossing needs fresh evidence under its own geometry and workload.
FAQ: forklift–pedestrian AI camera deployment
Can we reuse existing CCTV cameras for forklift proximity detection?
Possibly. First prove that their actual stream sees people and trucks early enough under representative loads, compression, latency, lighting and maintenance conditions. If a pallet hides a person until the crossing, a different model cannot recover unseen information. A route change may help more than another camera.
Can an AI warning replace a barrier or convex mirror?
Do not remove an existing measure solely because a camera has been installed. A barrier reduces contact opportunities physically; the camera warns only about conditions it can observe. Any removal needs a fresh site risk assessment and competent evaluation. OSHA and HSE both emphasize separation and visibility. OSHA pedestrian traffic · HSE route separation
What should a forklift safety-system acceptance test pass?
Test warning in each defined hazardous approach, actual recipient recognition, nuisance-alert burden, occlusion, loss of observation, fallback and recovery. Derive limits from that crossing’s stopping opportunity. A camera’s person-detection accuracy alone is insufficient; record vehicle and pedestrian responses.
Can it stop the forklift automatically?
Connectivity is not proof of a safe stop. Competent specialists must engineer and validate the complete function, including vehicle dynamics, communication failure, latency, false stops and restart. The advisory warning tests here cannot be reused as certification of automatic braking.
Who should join the first factory survey in Thailand?
Bring the operations manager, EHS, drivers, pedestrian-side supervisors, maintenance, IT/OT and the privacy or labor owner. A map of one crossing, existing controls and near-miss observations is enough to start a scoped discussion.
Conclusion: success is an early warning that people can act on
Reduce crossings first. Observe the residual blind corner far enough upstream, deliver an intelligible alert, and make the driver and pedestrian response explicit. Include occlusion, false alerts and faults in acceptance testing. Judge the system by performance in hazardous scenarios, not a single model score. Prove one crossing, then redesign for the next.
If your Thailand plant is still deciding which blind corner to survey or whether existing cameras provide enough warning time, contact TOMAS TECH with a route sketch and the current traffic rules. We can discuss feasibility before you commit to equipment.
