HIRI Visual AI Anti-Lifting: Warning triggered before the tire leaves the ground.

01 System Overview and Technical Background

During loading and unloading operations at the Quay, “ Anti-lifting ” What needs to be addressed is not a simple vehicle identification issue, but an interlock judgment that directly affects lifting safety: when a Quay Crane or Yard Crane is lifting a container, if the locking pin has not been fully released, one side remains locked, or the frame and the container are still mechanically coupled, the system must complete the identification before the entire vehicle or one half is lifted off the ground and reliably transmit the result. PLC Or a safety interlock, which triggers shutdown, soft start, or an alarm.

If this window is missed, an accident could occur.

Followed

Follow

Replay Share Like

Close

Watch more

More

Exit Full Screen

Switch to portrait mode, enter full screen, exit full screen.

HIRI Innovation has been followed.

Share video

0/0

00:00/00:09

Switch to landscape mode

Continue playing

From lasers to vision, why does anti‑lifting require AI?

As the level of port automation increases, what the system needs is no longer just “ See anomaly ” , but rather to stably distinguish within an extremely short lifting window “ Normal detachment ” and “ Abnormal belt lift-up ”。

Traditional solutions often rely on manual video inspection, edge displacement thresholds, or single-point laser ranging to determine whether an anomaly has occurred. These methods can function under stable operating conditions, but once they encounter real-world scenarios such as nighttime, rain and fog, backlighting, dust, tire rebound, or subtle Spreader vibrations, the risks of false alarms and missed detections increase significantly.

▲ Complex scenarios in real-world tasks

Take laser rangefinding as an example: In conditions of dense fog, heavy rain, or high particulate matter, the laser beam is scattered and absorbed by suspended particles, leading to attenuation of the return signal or even its complete loss. As a result, the system either fails to measure distance or outputs erroneous distance readings. Furthermore, lasers can only acquire distance information at one or a few discrete points, making it difficult to detect complex pose changes such as tire deformation or chassis tilt—directly undermining the safety benefits of anti‑lifting protection.

In contrast, visual AI Not only can it maintain operability in harsh environments through algorithmic enhancements—such as dehazing, rain removal, and low-light enhancement—but it can also simultaneously capture multi-dimensional evidence, including wheel‑ground contact patterns, vehicle frame posture, and relative motion of the cargo body, thereby providing a more comprehensive decision‑making foundation for lift‑prevention assessments.

 

Definition of the Visual Anti-Lifting System

The visual anti-lift system can be understood as an edge‑based safety perception system that relies primarily on industrial cameras, supplemented by depth sensors or laser scanners. It continuously monitors the contact between the tires and the ground, the posture of the vehicle’s underbody, the relative geometry between the cargo floor and the chassis, as well as short‑term upward motion trends, while integrating these inputs with a deterministic state machine and… PLC Interlocking, provide “ Safe release ”“ Suspicious slow start ” or “ Alarm shutdown ” The final action.

The Intelligent Trailer Lifting Prevent System launched by HIRI FUTURE ( Trailer Lifting Prevent System ), Adopting an end-to-end Artificial Intelligence model , without relying on manually set thresholds or logical judgments, That is, by training and constructing it using massive amounts of data. AI Model, It can perform real-time comprehensive analysis of complex conditions, including the tilt angles of the vehicle, wheels, frame, and containers. , comprehensively assess the risk of the container truck being lifted, and promptly and appropriately intervene to restrict equipment movements; currently, HIRI TLPS The system achieves the industry’s lowest lifting height, enabling operation in most scenarios. “ Tire never leaves the ground ” Early warning at the level, Coverage 70 Other typical lifting conditions 。

HIRI TLPS Core values:

· Move anomaly detection to before the vehicle actually leaves the ground, thereby achieving… “ Tire never leaves the ground ” Early warning at the level.

· Through end-to-end vision AI The model comprehensively captures the contact relationships among the wheels, frame, and housing with the ground, thereby reducing reliance on manual thresholds and single-point rules.

· By leveraging multi-frame temporal correlation and multi-evidence fusion, the risk of false alarms caused by normal lifting operations, equipment vibrations, and lighting variations is reduced.

· Edge-based real-time inference and interconnection PLC A closed-loop safety system enables rapid, reliable, and traceable handling of hoisting‑prevention measures.

02 Perception Link of the Anti-Lifting System

2.1 Industry Challenges: Why is anti-lifting detection more difficult than conventional inspection?

The challenge of preventing hijacking lies in the fact that it is not a static identification task, but rather a dynamic security assessment conducted on the millisecond to second timescale. What’s truly hard to tell apart is “ The housing has detached normally and is moving upward. ” , or still “ The container is lifting the reach stacker along with it. ”。 During this process, tire compression and rebound, elastic deformation of the vehicle frame, partial occlusion, lens contamination, glare from night‑time fill lighting, camera shake, rolling‑shutter distortion, and even image freezing or timestamp desynchronization can all cause the system to draw incorrect conclusions within an extremely short time window.

2.2 HIRI’s Core Technological Path: End-to-End Understanding Raises Risks

Tradition AI Visual solutions typically employ “ Detection target — Measure distance — Manually set threshold — Trigger alarm ” of the segmented Procedure. This Procedure can function under standard conditions, but at real Quay sites, factors such as tire rebound, frame elastic deformation, Spreader micro‑vibrations, obstructions, rain and fog, backlighting, and lens contamination make it difficult for a single threshold to remain stable over the long term. If the threshold is set too strictly, frequent false stops are likely; if it’s set too loosely, early lifting risks may be missed.

HIRI TLPS Adopt end-to-end vision AI The model consolidates the decision-making process, which was previously distributed across multiple rule-based modules, into a unified framework for comprehensive learning. The model takes as input not only the target bounding box in a single frame, but also multiple types of information across consecutive frames, including wheel position, frame geometry, cargo‑box configuration, ground‑contact area, tilt angle, and relative motion trends. By training on a large dataset comprising normal lifting, abnormal lifting, partial‑load lifting, single‑corner lifting, unlatched pin conditions, tire rebound, and other scenarios, The system is capable of learning. “ Normal detachment ” and “ Abnormal belt lift-up ” The subtle differences between… 。

 

In actual operation, End-to-end models perform integrated analysis of visual evidence within continuous short time windows, directly outputting a risk level or risk score for container‑truck lifting operations. Compared with traditional approaches that rely on human expertise to configure thresholds for each individual parameter, this model demonstrates superior adaptability to complex operating conditions and offers distinct advantages in false‑alarm suppression, early warning, and cross‑scenario generalization.

It is important to emphasize that end-to-end does not imply the elimination of safety interlocks. The model is responsible for identifying risks more accurately and at an earlier stage, while the final control actions—such as slow start-up, shutdown, and alarm activation—remain governed by deterministic state machines and… PLC Interlocking is executed. In this way, it both leverages… AI The model’s capacity to understand complex visual states, while retaining the controllability, auditability, and determinism that industrial safety systems must possess.

2.3 HIRI's Engineering Plan

To realize the aforementioned perception link, HIRI TLPS The following four key design features were adopted in the engineering:

1. Dual-channel synchronous industrial RGB Camera as the primary sensor, with depth/laser as supplementary.
The system deploys two synchronously operating industrial RGB cameras on the crane’s side beams, with one focused on the tire–ground contact zone and the other on the area adjacent to the bottom of the frame box. This dual‑camera configuration enables simultaneous capture of both local contact details—such as tire deformation and wheel‑ground clearance—and global geometric changes—such as left–right attitude differences and frame lift. Depth sensors or sparse LiDAR are not used as primary perception modalities; they serve only to perform secondary verification of candidate anomaly regions when visual inputs are obstructed or confidence levels are low. This “vision‑first, depth‑second” architecture reduces hardware costs and system complexity while maintaining high accuracy. 2. Key ROI + anchor-free Detection + Lightweight Tracking

The model tier does not directly process the entire image; instead, it first delineates regions of interest that contain the tire, the lower edge of the frame, and the bottom boundary of the box ( Return on Investment ). In Return on Investment Internally adopted anchorfree The detector localizes key regions, and this detection approach does not rely on predefined anchor boxes, features few hyperparameters, is well-suited for edge deployment, and exhibits strong robustness to object deformation. Subsequently, a lightweight multi-object tracking module maintains stable identity associations across consecutive frames for the same tire and the same vehicle frame. Even in cases of partial occlusion or brief low-confidence detections, the tracker preserves target continuity, preventing misjudgments caused by missed detections in a single frame.

3. Joint Modeling of Optical Flow, Depth Residuals, and Geometric Metrics

The greatest advantage of the anti-lifting system is that it does not rely on a single visual feature; instead, it simultaneously extracts three types of information:

· Optical Flow and Motion Field: Calculate pixel-level motion direction and velocity to determine whether the frame or tires are experiencing abnormal upward movement.

· Geometric measurement: Explicit dimensions such as wheel‑ground clearance, the distance from the lower edge of the frame to the ground, and the left–right attitude difference are measured from the image.

· Deep Residual (When the depth sensor is engaged): Verify whether the spatial clearance between the box bottom and the vehicle frame actually exists.

The three types of evidence cross-validate one another: Detection and tracking provided “ Who’s moving? ” , optical flow provides “ How to move ” , geometric quantities are provided “ How much did it move? ” , provide in depth “ Are you still in contact? ” Joint modeling is more robust to interference than relying solely on the positional changes of bounding boxes, and it better distinguishes between normal disengagement and abnormal lifting.

4. Short-term sequence fusion + deterministic state machine + PLC Interlocking

The perception layer does not output a single-frame classification result, but rather a sequence over a short time window (for example 10 ~ 15  the risk score within the frame. This score is then fed into a deterministic state machine—i.e., a rule-based module—which, based on predefined threshold values, consecutive-frame confirmation, the current lifting phase, and other rules, determines the output. “ Safe release ”“ Suspicious slow start ” or “ Alarm shutdown ” . This tiered approach is easier to audit and better aligns with the implementation logic of industrial security systems.

03 Beyond recognition—reliable shutdown is what truly matters.

Whole vehicle, half-side, and single-corner anomalies are uniformly identified.

HIRI TLPS It not only triggers an alarm when the entire vehicle is lifted, but also provides early, more subtle warnings during earlier, less obvious stages—such as left–right asymmetry in posture, a single-side tire lifting off the ground first, or an abnormal elevation of a single corner. For on-site operations, this… “ Lead time ” It is often more valuable than simply lowering the ground clearance threshold. —— Because truly dangerous accidents typically begin with small, asymmetrical anomalies that gradually escalate.

Multi-evidence consistency effectively prevents false shutdowns.

HIRI TLPS By jointly validating multiple types of evidence—such as geometric relationships, lifting speed, depth residuals, the hoisting encoder, and temporal continuity—it is possible to clearly distinguish between normal lifting, Spreader micro-vibrations, tire rebound, and genuine anomalies, thereby avoiding frequent false stops that would otherwise reduce operational efficiency.

Health diagnostics and tamper-proofing, implemented as a standalone blockchain.

The system has independent health‑diagnosis capacity and can detect in real time conditions such as shading, defocusing, lens tampering, video freezing, and timestamp anomalies. Once the camera’s health status becomes distorted, the system automatically switches to a conservative mode—such as allowing only slow start-up, requiring manual confirmation, or outright rejecting lift-off—to ensure that safety thresholds are not breached.

Suggested key verification scenarios to cover:

· Various combinations of lighting and weather conditions, including daytime, nighttime, rain and fog, backlighting, and strong glare.

· Engineering disturbances such as lens contamination, partial occlusion, slight vibration of the mounting bracket, and edge distortion.

· Key samples such as full-vehicle lifting, half-side lifting, single-corner lifting, and rapid ascent following normal release.

· Failure and tampering scenarios such as frame loss, freezing, timestamp anomalies, and single-sensor distortion.

 

04 Our Key Optimization Directions

Supplement complex boundary samples

Key improvements include semi‑side lifting, single‑corner lifting, partial release of the locking pin, and frame rebound with normal disengagement. “ The hardest to distinguish ” The capacity for scene recognition moves risk assessment forward to the stage when anomalies are just emerging.

Enhance all-weather robustness and self-diagnostic capacity

Nighttime supplemental lighting, rain and fog, backlighting, dirty lenses, obstructions, vibrations, frozen frames, timestamp anomalies, and similar issues must be addressed in tandem with environmental adaptability assessments and equipment health monitoring to prevent… “ Can't see clearly ” And lose one’s judgment.

Optimizing Edge Deployment Efficiency and Interlocking Strategies

By triggering deep confirmation within the candidate window, we reduce unnecessary computations, shorten alert latency, refine tiered response actions for slow startups and hard stops, and enhance event logging and state transition records, thereby achieving more stable, faster, and more auditable anomaly handling.

Enhance the generalized capacity across vehicle models, cranes, and port areas.

Continuously accumulate data on various devices, frames, container types, and operating conditions, and refine the current solution from… “ Available for a single project ” Iteration is “ Replicable across multiple scenarios ” The standardization capability module reduces subsequent deployment and maintenance costs.