Forklift-Pedestrian Safety AI Checklist
Start with the conflict point
Forklift-pedestrian safety is a field problem before it is a software problem. The first question is not which model to run. The first question is where people and vehicles actually cross paths. Walk the site and mark the specific blind corners, dock lanes, charging areas, freezer entrances, and staging zones where a warning would matter.
Checklist 1: camera readiness
Confirm that the current camera sees the floor area where the conflict happens. A high camera that only shows the top of racking may be useful for security review, but weak for path analysis. A useful view shows the vehicle path, the pedestrian path, and the seconds before intersection.
Stream access
Verify that the camera exposes a stable RTSP or ONVIF stream and that the stream remains stable during normal shift load. Do not test only after hours. Forklift safety systems have to work when the network and floor are busy.
Lighting and occlusion
Check glare, shadows, low light, rack occlusion, door openings, and floor markings. The goal is not a perfect camera. The goal is a camera that sees enough of the conflict to trigger a reliable warning.
Checklist 2: alert path
A forklift warning can go to several places: the VMS, a supervisor tablet, a local horn, a light stack, or a control room. Pick one primary path for the pilot. If everyone receives every alert, no one owns it.
Local signal or supervisor review
If the goal is immediate intervention, a local signal near the conflict point may be required. If the goal is retraining and hot spot analysis, supervisor review may be enough for the first phase. The system should match the action you expect.
Checklist 3: model criteria
Do not evaluate forklift safety AI only on whether it detects a forklift and a person. That is the easy part. Evaluate whether it understands path, speed, distance, occlusion, and time to conflict. A stationary forklift beside a pedestrian is not the same as a moving forklift entering a blind corner.
Near miss versus collision risk
Define what counts as a near miss before the pilot. Sudden stops, close passing, path intersection, and unsafe speed near a walking path may all be useful categories, but they should not be mixed without labels.
Checklist 4: pilot metrics
Track alert latency, event volume, supervisor agreement, repeat hot spots, and whether the clips are useful in a safety meeting. A good pilot should produce fewer arguments and better evidence. It should help the team see where the risk repeats.
Checklist 5: privacy and retention
Raw video should remain in the current VMS unless policy says otherwise. DHI can send structured event data while the footage stays on site. That matters for warehouse operators who need safety evidence without creating a new unmanaged video store.
Final decision
At the end of the pilot, the decision should be concrete. Did the system find real conflicts? Did supervisors trust the clips? Did the alert arrive in time? Did a pattern emerge by aisle, dock, or shift? If the answer is yes, the next step is not another demo. It is a zone-by-zone expansion plan.