Pedestrian detection on excavators and wheel loaders: what the camera is documented to do, and what no datasheet says

The task is the same on every site: behind a wheel loader or beside an excavator stands a person the driver cannot see from the cab, and the machine should know that before the driver does. Two documented routes lead there. The first is a retrofit. Blaxtair, a product of the French manufacturer Arcure, is a camera with embedded artificial intelligence that the maker has had in the field since 2009 and that, by its own account, is compatible with all makes, sizes and types of excavators, wheeled or tracked, and likewise with wheel loaders and forklifts. The second route is a factory option. Liebherr offers active personnel detection for its mid-sized wheel loaders and extended it in April 2021 with a brake assistant and a hazard map; the trade press reported on it on 7 April 2021.

Both routes solve the same task with the same basic idea, a camera whose analysis distinguishes persons from objects, and they differ in almost everything that matters for procurement: whether the machine is already in the yard or still to be ordered, whether the reaction is a warning or a braking intervention, and where the data goes. This article reads both manufacturers' product pages as of September 2026, and only those.

Blaxtair's product pages call the camera the only on-board AI system that accurately detects and locates pedestrians in real time; that is a manufacturer's claim and is reported here as one. Documented is detection in any posture, standing, crouching or partly visible, forwards as well as in reverse. Processing runs on board; a cloud is not required for detection itself. The sensor is rated IP69K with an operating temperature range of minus 40 to plus 85 degrees Celsius, designed for strong vibration and impacts. The maker describes installation as plug-and-play, compact and simple. Around pedestrian detection, Arcure offers further modules under the same name: detection of specific obstacles, vehicle detection, shock detection, a zoning function, a checklist function, a perception technology for original equipment manufacturers, and a cloud-based IoT connection called Blaxtair Connect, which the maker says divides the coexistence of vehicles and pedestrians by three. That too is a manufacturer's figure without a published method.

That is an honest list, and it is honest also in what it leaves out. On the pages read there is no detection range in metres, no reaction time in milliseconds, no statement on detection performance in darkness, rain or backlight, no false-alarm rate, no statement on whether and for how long images are stored, and no standard or certification for the function. A system designed to warn the driver need not state any of that; an operator who procures it still needs to know.

Liebherr's active personnel detection warns the operator, according to the maker, visually on the display and acoustically of hazards in the rear area of the wheel loader. To warn specifically of injury to persons, the assistance system automatically distinguishes between persons and objects, using camera technology and without persons having to carry transponders or transmitters; for persons it warns at a greater distance than for static obstacles such as walls or pillars. The brake assistant added in 2021 automatically slows the machine as soon as personnel detection identifies a hazard in the rear area, down to a standstill, and so intervenes earlier than a driver whose reaction time is removed from the chain. The second extension is the hazard map: every time the system detects a hazard and warns, it transmits a GPS point to Liebherr's data transmission and location system LiDAT, and the operator receives from that a map of those points showing where on the site the warnings cluster.

This description too leaves out what an acceptance would need: no distance at which braking begins, no deceleration in metres per second squared, no performance level or category to ISO 13849 for the braking function, no statement on whether the machine reacts differently to a warning in forward travel, and no statement on whether images are stored or only event points transmitted. The hazard map is the detail that raises the data question: one GPS point per warning is an event without an image, and whether it stays that way the description does not say. Whether the function is standard equipment or an option, and whether it can be retrofitted to older machines of the series, is likewise not clear from the coverage read; that is a question for the dealer.

Set the two descriptions side by side and a list emerges of what is missing for the decision and can only be had on site or in the contract. Range: from what distance a person is detected, standing and crouching, and from what distance the system warns or brakes. Time: how long passes between detection and warning, and with the factory option between detection and the start of braking. Light: what is detected at night, in low sun, with rain on the lens and in dust; the IP69K rating says the sensor survives that, not that it then sees. False alarms: how often the system triggers on puddles, traffic cones, fence panels or a discarded overall, because a driver warned ten times a day for nothing stops listening after a week. And safety level: for a function that brakes a machine on its own, the question of the performance level to ISO 13849-1 is not a formality but the difference between a comfort function and a safety function, and none of the pages read answers it.

The route to those numbers is the same as for any machine with sensing: a demonstration on the operator's own site, in its own light, with its own load carriers and an employee in its own high-visibility clothing, crouching behind the rear; and a contract that records the values the demonstration produces as warranted properties. A datasheet that does not state the numbers is not a failing of the manufacturer. It is an instruction to measure them.

A camera that detects people is not a camera that identifies people, and that difference carries the decision whether such a system on a construction site is a data-protection matter or not. Both manufacturers describe their systems as on-board detection: the analysis establishes that a person is in the danger zone, not who. As long as nothing is stored and nothing transmitted, it stays that way. The case changes with every add-on that sends events outward: Blaxtair Connect is described as a cloud-based IoT solution, and Liebherr's hazard map transmits one GPS point per warning into LiDAT. Whether an image goes with it, a crop, only a coordinate or only a counter, none of the sources read says, and that is exactly the question that goes to the supplier before the order and belongs in the contract.

The answer decides who in the company has to be involved. A system that detects on board and only warns affects the employees on the site no differently from a reversing alarm. A system that transmits events with images to a cloud captures employees at their workplace, and the assessment then belongs to the data protection officer and, where there is one, to the works council, before the first machine drives with it. This article does not make that assessment; it records that it depends on a contract clause the datasheets leave open.

Both systems share the same architecture, and it is worth naming because it decides the evidence. The detection part is learned: a model decides from camera images whether the object behind the machine is a person. The reaction is deterministic: warning tone, display, and with Liebherr a braking sequence by fixed rules. For the learned part there is no enumerable list of cases to accept; one can only cover scenarios, crouching, half hidden, in backlight, and document the coverage. For the deterministic part the rules of machinery safety apply, and for a braking intervention that means the function itself needs evidence that does not follow from the camera's detection rate.

That is not a peculiarity of these two products. It is the split that applies to every learned component on a machine, and it is the reason a personnel detection that detects well is not yet a safety function that brakes well. Anyone buying both in one purchase is buying two pieces of evidence, and the second is in none of the product descriptions read.

For an existing fleet the retrofit is the route, and the question is not whether it works but which numbers the demonstration on the operator's own site produces and whether they are in the contract. For a new purchase the factory option is the route if the manufacturer offers it for the series in question, and the question is then whether the braking function is evidenced as a safety function or sold as assistance. In both cases the question about the images goes before the order. werob is a manufacturer-independent systems integrator; Arcure and Liebherr appear here as publicly documented market examples and not as confirmed partners, and werob has neither accepted nor measured either system. What werob records in a specification for a site are the scenarios the demonstration has to cover and the values the contract should warrant; the measurement itself takes place on the operator's site.

The Blaxtair details come from the product pages of the manufacturer Arcure (category pages for excavators and loaders, product page for pedestrian detection), read in September 2026; manufacturer claims such as the only system of its kind or the division by three are marked as such and unverified. The Liebherr details come from trade-press coverage of 7 April 2021 of the brake assistant and hazard map for the mid-sized wheel loader series; series designations, equipment status and retrofittability were not verified with the manufacturer. None of the sources gives range, reaction time, false-alarm rate, image storage or a performance level; where this article names those quantities it names them as missing, not as measured. No measurement, acceptance or customer project of werob's underlies this article, and no figure on this page is werob's own.