Pedestrians struck by vehicles, particularly by handling equipment, are among the leading causes of death in cargo handling. At Busan New Container Terminal, CyberLogitec has combined AI vision, RTK positioning and a real-time digital twin so that workers and straddle carriers share a single map, and has built its alerting logic around one requirement: drivers must be able to trust it.

In August 2026, PEMA, TT Club and ICHCA International jointly called for stronger collision-prevention measures across ports and terminals. They noted that vehicles, heavy machinery and personnel often work in close proximity, and that advanced non-contact collision-prevention technologies are not yet covered by national or international standards.^[1]

The case for action is well established. ICHCA International’s Severe Risks Dashboard, which analyses more than 500 cargo-handling fatalities recorded since 2000, identifies pedestrians run over by vehicles, particularly by handling equipment, as a close second among the causes of death. It describes the effective segregation of people and vehicles as “a safety bedrock”.^[2]

In South Korea, the pressure is also legal. Since the Serious Accidents Punishment Act took effect in 2022, senior executives can face criminal liability for fatal workplace accidents. For terminal operators, collision prevention is no longer only a safety goal. It is a board-level responsibility.

Yet segregation has limits. Beneath a ship-to-shore (STS) crane, lashers, checkers and straddle carriers have to share the same ground. This is the setting in which we worked with Busan New Container Terminal (“BNCT”), the operator of Terminal 5 at Busan New Port and a straddle carrier-operated facility, to build a system that answers a deceptively simple question: how far apart are this machine and this person, right now?

Why tags and cameras alone fall short

Proximity detection is not new. Tag-based systems, in which workers wear radio transponders that equipment can sense, are widely used in mining and in some terminals. They work well when everyone wears a working tag. But tags are forgotten, lost or left with flat batteries, and an untagged worker is invisible to the system.

CCTV has the opposite limitation. It sees everyone, but it produces pictures, not positions. An operator watching a monitor cannot measure distance, and cannot watch every camera at once.

At the root of both problems is the same gap. Equipment exists as satellite coordinates. People exist as pixels. Until the two are expressed in the same frame of reference, distance cannot be computed reliably.

A shared map as the foundation

The foundation of the BNCT system is CyberLogitec’s digital twin-based real-time port monitoring platform. It replicates the physical terminal in a virtual space and updates the movement of cranes, carriers and people in real time. Crucially, it places people and equipment on a single coordinate system.

To bring people onto that map, we worked with GlobeAI to develop a vision-to-geolocation technique, for which the two companies have filed a joint patent application. AI cameras mounted on the STS cranes detect workers, and each detection is converted from image coordinates into real-world UTM coordinates using the crane’s known position and the camera’s geometry. Straddle carriers fitted with RTK receivers report their positions into the same frame. The platform is integrated with the terminal operating system, OPUS Terminal, so that safety data sits alongside operational data (Figure 1).

Once a worker and a carrier appear on the same map, the distance between them becomes a number the system can compute and act on. Workers need no tag or device.

Figure 1. System architecture at BNCT. AI cameras on STS cranes and RTK receivers on straddle carriers feed an IoT platform that fuses both onto a single digital twin map, triggering real-time alerts in the control room and in the driver’s cab. (Illustration: CyberLogitec)

Engineering for trust

A safety system can fail in two ways. A false negative, where a real risk goes unflagged, can be fatal. A false positive, where an alarm sounds when no one is in danger, is less visible but just as corrosive.

The lanes beneath an STS crane are narrow. A colleague working safely in the adjacent lane is often only a few metres from a passing carrier. A system that looks at distance alone would alarm constantly. When drivers hear dozens of meaningless alerts a shift, they learn to ignore them, or switch the system off, and then it fails exactly when it is needed.

Tag-free detection addresses false negatives: anyone the cameras can see is covered. For false positives, we used the spatial model already in the digital twin.

The berth at BNCT has four working lanes beneath the STS cranes, two lanes in the back-reach area and safety zones between them. We defined each as a non-overlapping geofence, so that every carrier and every worker belongs to exactly one zone at any moment. In effect, the geofences apply the principle of segregation digitally, in the one area where physical segregation is not possible. The system then evaluates risk in a fixed sequence:

1.   Identify the carrier’s lane. If the carrier is crossing a safety zone, no evaluation is made.

2.   Identify each worker’s zone.

3.   Compare them. If the zones differ, the alert is suppressed, regardless of distance.

4.   Only if they share a lane is distance calculated and an alert level assigned.

The sequence runs on every detection cycle. A worker who steps from a safety zone into a carrier’s lane triggers a new alert immediately; one who steps out clears it.

Alerts are tiered at 15 metres (caution), 10 metres (warning) and 5 metres (danger). On the carrier’s vehicle-mounted computer, they appear as directional bars at the front, rear, left or right of the screen, so the driver knows at a glance where the risk lies.

Given that the industry has yet to standardise non-contact collision prevention, we would argue that alert credibility should be a design criterion from the outset, not a tuning exercise afterwards. A system that drivers switch off protects no one.

Specialist technologies on a common platform

A container terminal is one of the hardest environments for both computer vision and satellite positioning. We integrated specialist partners on top of the digital twin.

Worker detection — GlobeAI. GlobeAI’s deep-learning models detect workers against cluttered container backgrounds, through day–night lighting changes, sea fog and rain, and when workers are only partly visible behind equipment. The models were refined using real port imagery. Because every coordinate we generate starts with a detection, this performance sets the reliability ceiling for the whole system. The conversion of those detections into real-world coordinates is the subject of a joint patent application by CyberLogitec and GlobeAI.

Equipment positioning — LG Uplus. Steel container stacks reflect satellite signals and crane structures block them, producing multipath errors that defeat conventional GNSS. LG Uplus’s RTK network provides continuous correction data from more than 200 ground reference stations across Korea, allowing straddle carriers to be tracked to centimetre accuracy.

At BNCT, RTK receivers are installed on around 70 mobile units, and 28 AI cameras are mounted on 14 STS cranes, two per crane.

Privacy by design

Camera-based systems raise legitimate questions from workers and unions. The BNCT system is designed to determine where people are, not who they are. It detects the presence and position of a person for the sole purpose of collision prevention and does not identify individuals.

From investigation to prevention

In the control room, the data is rendered in CyberLogitec’s Virtual Terminal, a 3D digital twin of the site. Operators switch between 2D and 3D views, split the screen to watch several areas at once, and follow a selected machine as it moves. Hazard zones can be drawn directly onto the map.

A rewind function allows supervisors to replay any moment and see exactly where people and machines were when a near-miss occurred. Every alert is stored as an auditable record, and a management dashboard tracks weekly reports and a terminal safety score. The emphasis shifts from investigating incidents after the fact to redesigning workflows before they happen.

Han-Seo Moh, Senior Vice President and Head of CyberLogitec’s Port Solution Group, says: “The hard part of port safety was never tracking equipment. It was locating a worker who carries no device, against the same reference as a machine moving with centimetre precision. Our answer was to give people coordinates.”

“How many boxes a terminal moves in a day matters less than whether everyone who clocked in goes home safely,” he adds.

Where things stand

In field validation at BNCT, RTK equipment positioning achieved an accuracy of under 3 centimetres, and AI vision worker localisation was accurate to under 20 centimetres. The system is now in operation across BNCT’s berth area.

“Safety is never open to compromise and remains our highest priority,” says Chris Choi, CEO and Representative Director of BNCT. “Through our collaboration with CyberLogitec, we have further strengthened our prevention-focused approach to safety management. We will continue to proactively manage operational risks and provide a safer working environment for everyone working at the terminal.”

Lessons for other terminals

Several principles from Busan apply well beyond a single site:

•     Solve the geometry first. Collision prevention is fundamentally a spatial problem. Until people and equipment share a coordinate frame, more sensors will not answer the basic question of distance.

•     Segregate digitally where you cannot segregate physically. Geofenced zones in a digital twin extend the logic of physical separation into areas where people and machines must share space.

•     Cover everyone, not only the tagged. Vision-based localisation closes the gap left by missing or failed wearables, and can complement tag-based systems where they already exist.

•     Treat trust as a requirement. Context-aware logic such as lane matching matters as much as raw accuracy.

•     Integrate specialists on a neutral platform. No single supplier leads in vision, positioning and terminal operations. A digital twin gives them common ground, and the same architecture can be extended to terminal tractors and other mobile equipment.

As the industry works towards common standards for non-contact collision prevention, we believe field experience of this kind has a useful part to play.

References

[1] SAFETY4SEA, “Industry groups urge stronger collision prevention measures across ports & terminals”, 13 August 2026 (PEMA, TT Club and ICHCA International).

[2] ICHCA International, Severe Risks Dashboard (fatality data 2000 to July 2024; last updated 5 June 2025). https://ichca.com/ichca-severe-risks-dashboard

Author: Dr Junhee Cho, Digital Team Leader at CyberLogitec, where he leads the development of AI and digital twin solutions for port operations. He holds a PhD in computer vision and is an adjunct professor at Soongsil University, Seoul.

The article or images cannot be reproduced, copied, shared or used in any form without the permission of the author and CyberLogitec.

Copyright,2026, All rights reserved