The blind corner where a rack aisle opens onto the main aisle is where forklifts and pedestrians come face to face most often in warehouses; a mirror is fitted, the floor is marked, and the risk still remains. At this point, video analytics can detect the forklift and the person separately in the camera image and generate a warning based on the distance between them or on both entering a defined hazard zone together. But it’s useful only if it’s installed with the right expectations. Below we explain the use scenarios, zone and proximity definition, camera position, latency budget, warning mechanisms and why analytics can’t replace a safety system.
Use cases
Forklift–pedestrian analytics isn’t a single rule but the common name for several different scenarios. The target scenario determines camera and rule design:
- Blind corner and intersection warning: At points where rack aisles open onto the main aisle, warning the pedestrian on the other side as a forklift approaches. The warning is usually a light sign or audible alert at the corner.
- Pedestrian exclusion zone: Pedestrian entry into areas where only forklifts operate (loading area, block stacking area).
- Proximity in shared areas: In places where forklifts and pedestrians work in the same area, the two coming closer than a certain distance.
- Dock and door opening: A pedestrian being present at truck loading doors while a forklift operates between the dock and the vehicle.
- Reporting and risk map: Instead of instant warnings, reporting at which points and times near misses concentrate.
Unauthorized entry into a static zone is a single-object virtual fence problem and is set up with the method in the virtual fence and area intrusion rule design article. Here, the subject is the position of two moving objects relative to each other. That’s where the difficulty comes from.
Zone and proximity definition
Proximity is defined in several ways; systems usually support one or two of them:
| Method | How it works | Prerequisite |
|---|---|---|
| Zone co-occupancy | An event if the forklift and pedestrian are in the same defined zone at the same time | A correctly drawn ground zone |
| Pixel distance | A threshold on the image distance between the two objects’ boxes | Accounting for perspective |
| Calibrated metric distance | The image is calibrated to the ground plane, and distance is calculated in meters | Camera calibration, flat floor |
| Dynamic zone around the forklift | A zone around the forklift that grows in the direction of travel | Direction and speed estimation |
Zone co-occupancy is the simplest and most predictable method. At an intersection, the aisle the forklift comes from and the aisle the pedestrian comes from are defined as separate zones; if there’s an object in both at once, a warning is given. No distance calculation is needed, and the setup is easy to test. This is the first choice on site.
Pixel distance is misleading because of perspective: 200 pixels between two objects near the camera equals a few meters, while the same 200 pixels far away corresponds to a much larger distance. This method only makes sense with cameras looking from above.
Calibrated metric distance gives the most meaningful result, but it requires camera calibration and a flat floor. Under racks, on ramps and on sloping floors, the error grows. If you use this method, verify the calibration with known distance markings on the floor.
Whatever the method, choose the proximity threshold according to the forklift’s speed and stopping distance. The principle is: the warning must leave the pedestrian and operator enough time to react. This ties directly to the latency budget.
Camera position and coverage
In enterprise camera systems, camera position for forklift–pedestrian analytics is chosen by different criteria from general warehouse surveillance:
- High and steep angle: In rack aisles and intersections, placing the camera high, looking from above or at a steep angle, reduces the forklift hiding the pedestrian and loaded forks blocking the view.
- Covering the approach distance: The camera must see not only the intersection but also the parts of both aisles approaching it. If the forklift isn’t detected before it reaches the intersection, the warning comes too late. A late warning isn’t a warning.
- Loads and reflections: Stretch-wrapped pallets, metal racks and wet floors produce reflections. Keep rule zones away from these surfaces.
- Motion blur: A fast forklift looks blurred at long exposures, and classification gets harder. Limit the maximum shutter time on analytics cameras; details are in the camera exposure settings article. A mistake we often see on site is leaving the camera on default “auto” exposure; because it’s dark between the racks, the shutter lengthens and the forklift turns into a smeared blob.
- Lighting: Light levels between racks can be markedly lower than on the main aisle. If the camera switches to IR mode, forklift lights and reflectors create glare.
The warehouse and logistics camera design guide is a complementary resource for general camera design of a warehouse layout.
Latency and warning mechanisms
In proximity analytics, speed determines the value of the warning. Latency accumulates at every link in the chain: capturing the image, compression and network transfer (in a server-based architecture), decoding, running the model, the rule’s persistence time, the warning command reaching the output device and the person’s reaction. The calculation below is entirely illustrative; it shows the items you need to measure on your own system:
# Example latency budget (values are assumptions; measure on site)
capture_and_encode = 0.15 # s, camera side
network_and_decode = 0.10 # s, in a server-based architecture
inference = 0.05 # s, model run time (per frame)
analysis_fps = 10 # frames/s
persistence_frames = 3 # consecutive frames to confirm the event
output_device = 0.20 # s, I/O module or network relay
human_reaction = 1.00 # s, noticing the warning and acting
system = (capture_and_encode + network_and_decode + inference
+ persistence_frames / analysis_fps + output_device) # 0.80 s
total = system + human_reaction # 0.80 + 1.00 = 1.80 s
forklift_speed = 10 / 3.6 # 10 km/h ≈ 2.78 m/s
print(round(forklift_speed * total, 1), "m") # ≈ 5.0 m
In this example, the forklift travels about 5 meters while the warning is generated and the person reacts; braking distance is added on top. The warning zone must start ahead of this distance. This calculation describes an example operator warning scenario. It can’t be used on its own to choose a safety distance or a machine safety function. Two design decisions follow from the calculation: don’t keep the analysis frame rate and persistence time unnecessarily low, and generate the warning on site. Having the decision made on the camera or an edge device near the intersection reduces network and central processing latency; we compared architecture options in the edge or server article.
Warning mechanisms
- Light signs and beacons: At the intersection or the head of the aisle, in a position visible to both pedestrian and operator. A constantly flashing sign is ignored within a few weeks; it should activate only at the time of an event.
- Audible alerts: Limited effect in noisy areas, and a nuisance if triggered frequently.
- Floor projection or screen: Some facilities use a screen at the head of the aisle or a floor projection.
- Reporting: Each event is recorded in the VMS with a short clip; the OHS team reviews near-miss points weekly.
Relationship to safety systems, and limits
By its nature, video analytics is a probabilistic detection system: under certain conditions it can miss a person, be delayed by light and occlusion, and stop completely on a network or server failure. It isn’t designed as a machine safety function and shouldn’t be used on its own for safety decisions such as stopping a forklift. This line must be clear.
| Topic | Video analytics | Safety system (e.g. certified sensors) |
|---|---|---|
| Purpose | Awareness, warning, reporting | A function that prevents the hazardous situation |
| Failure behavior | Possibility of misses and false alarms | Transition to a defined safe state |
| Verification | Pilot and statistical testing | Against standards such as ISO 13849 |
| On failure | Can silently stop working | The fault is detected and reported |
| Its place | An additional layer of layered measures | Engineering control |
Machine safety-related control systems are assessed against standards such as the ISO 13849 series; the standards can be reviewed on iso.org. Video analytics shouldn’t be presented as a safety function unless it has passed this kind of assessment. Segregated pedestrian walkways, physical barriers, mirrors, speed limits and proximity sensors on the forklift stay in place; analytics is an awareness layer added on top of them.
Event reporting and improvement
The most lasting benefit of proximity events is making the risk map visible, more than the instant warning. Track these indicators in weekly or monthly reports:
- Number of near misses per intersection or zone, and their distribution by time.
- Relationship with periods such as shift changes and shipping peaks.
- Comparison before and after physical changes (mirrors, barriers, redrawn walkways).
- Changes in the number of false alarms; a sudden increase usually points to camera drift or a layout change.
Reports should target points and processes, not individual operators. Event clips are deleted after the set period, and only numerical data is kept for trend analysis. After every physical change, re-verify the relevant rule with a short test; at an intersection where racks have moved, the old zone is now in the wrong place.
For the design and pilot scope of hazard zone analytics, see our forklift and hazard zone analytics page; if PPE monitoring is planned in the same area, see our PPE detection and occupational safety video analytics page.
Frequently asked questions
Can forklift–pedestrian proximity analytics be used to stop the forklift?
It shouldn’t be used on its own. Video analytics works probabilistically; it can miss, be delayed and stop silently on failure. Safety functions such as stopping a machine should be provided by systems assessed against standards such as ISO 13849; analytics should remain an additional awareness layer.
How is the proximity threshold set?
The threshold is calculated from the forklift’s speed and the sum of system latency and human reaction time. The distance the forklift travels during this time, plus braking distance, gives the distance at which the warning must start. Verify the values by measuring on site.
Are existing warehouse cameras enough for this analytics?
Usually partly. The camera must also see the aisles approaching the intersection, be at a height where the forklift won’t hide the pedestrian, and have exposure settings that limit motion blur. Assess existing cameras against these criteria.
Should the warning be generated on the camera or the server?
For intersection warnings where latency is critical, the decision should be made on the camera or an edge device near the intersection, and the warning triggered via local I/O. The central server can be used for reporting and clip archiving.