In a warehouse, the accident happens with the forklift. The moment the pedestrian does not hear the forklift coming from behind, the moment the operator does not see the pedestrian behind a loaded fork. Catching that moment with a camera is possible; but "forklift and person in the same frame" does not mean accident. In a warehouse there is a forklift and a person in every frame. The problem is distance, direction and speed.
Metres, not pixels
What the model gives you is two boxes: forklift and person. The pixel distance between the centres of the two boxes means nothing; two objects near the camera look far apart, two objects far away look close. You need a homography against the floor: four points in the camera image are matched to four points on the floor, and the bottom-centre of each box is converted to floor coordinates. From then on distance is in metres.
For calibration we stick four measured tapes on the warehouse floor, mark them from the camera, done. If the camera moves (a rack collision, cleaning) the calibration breaks; for that there is a weekly automatic check, and if the reference points have shifted an alert goes out.
A velocity vector, not a single frame
Forklift parked, operator standing next to it: distance 1 metre, no danger. Forklift approaching at 8 km/h, pedestrian 3 metres ahead with her back turned: danger. The difference is speed and direction. Tracking data gives each object's floor coordinate over the last 10–15 frames, and from that comes a velocity vector. The rule takes this shape:
- Closing speed (the rate at which the distance between the two is shrinking) above a threshold
- Estimated time to intersection (when the distance reaches zero at the current speed) under 2 seconds
- The pedestrian's orientation is not towards the forklift (derived roughly from box orientation and movement direction)
A warning when all three hold. The first version, which looked only at distance, produced 400 warnings a day and nobody listened. With speed and time to intersection added it dropped to 20, all of them real approaches.
Graded warnings
A single type of alarm does not work; alarm fatigue sets in within a week. We built three grades:
| Grade | Condition | Action |
|---|---|---|
| Info | Distance under 3 m, approach slow | Logged only, for reporting |
| Warning | Time to intersection under 2 s | Light and sound beacon in the zone (IoT relay) |
| Incident | Distance under 1 m and speed not zero | Notification to shift supervisor, event frame stored |
The light-and-sound beacon is triggered through a relay on site; the server sends a message over MQTT and a small controller in the zone pulls the relay. This is where camera analytics meets IoT. Latency from server to relay is under 100 milliseconds, because everything is on the same local network.
Zone rules
The warehouse has the pedestrian walkway and the forklift corridor marked separately. In the analytics these lines are defined as polygons. A pedestrian staying in the forklift corridor for more than 5 seconds is one event; a forklift entering the pedestrian walkway is another. Both are independent of any accident; they measure behaviour. A line in the weekly report such as "walkway violations per shift went from 12 to 3" shows that the safety culture is changing. The customer's safety team takes that number to the management meeting.
Product-level video evidence
The logistics customer had another headache: the "goods arrived damaged" complaint. The camera at the loading dock, matched to the vehicle plate or the shipment number, stores the loading moment as a 30-second clip. When a complaint comes in they search by shipment number and watch. That is not pure video analytics, but it runs on the same infrastructure and for the customer the gain is at least as tangible as safety.
The operator side
In the first week operators saw the system as "they are watching us". Two things changed that: no identity, and frames stored only for a real incident; and the report is per zone, not per operator. The system answers "in which corridor, at what hour, does risk rise", not "who made a mistake". Set up that way, the shift supervisors became the first advocates, because they had numbers in hand to change the walkway layout.
The 96% vehicle-safety figure we saw in the field shows what share of defined proximity events were prevented before or by the beacon. That number varies with warehouse layout, forklift count and shift pattern; it is one site's number, not a sector average.