Automatic Intermodal Loading Unit Identification using Computer Vision: A Scoping Review

arXiv:2509.17707 · cs.CV · Submitted 2025-09-22 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Automatic Intermodal Loading Unit Identification using Computer Vision".

Paper summary: Tom: So we're looking at this scoping review about automatic intermodal loading unit identification using computer vision. Essentially, the paper maps out all the existing research on using computer vision for identifying these loading units in ports and terminals between nineteen ninety and twenty twenty-five.

Jane: That sounds really practical, Tom. What's the main point they’re trying to make about this whole area? Like, what's their core thesis here?

Tom: The paper’s main thrust is to survey the landscape of computer vision methods for identifying these intermodal loading units, and it tries to clarify all the different ways people have been doing it over those thirty-five years. It highlights where the field has come from and where it's currently stuck.

Lu: It starts by looking at how the technology has evolved, which is a really interesting way to frame the whole thing because you see a clear shift in approach.

Meng: What kind of evolution are they talking about? Is it just getting better software, or is there a change in what kind of hardware is used?

Tom: It’s both. They show this move from older, static systems like simple optical character recognition gates to more dynamic setups involving vehicle mounted cameras. This shift is key because it allows for much more precise monitoring of the units.

Jane: So, if we’re talking about precision, what kind of accuracy are we looking at in these identification methods? Is it perfect from the start?

Tom: The reported results for end-to-end accuracy range quite a bit, from five percent up to ninety-six percent across all these different studies. That wide range shows just how much the method depends on what you’re using and how good your data is.

Lu: And they track that evolution over time, which really helps show the historical development of the field. They point out a big turning point around two thousand sixteen to twenty twenty, where deep learning methods started to take over as the main approach.

Meng: Deep learning sounds promising, but what does that mean for putting something into a real port operation? Does it just make things faster, or does it introduce new kinds of problems when you try to deploy it on busy traffic?

Tom: Well, the paper suggests that moving towards deep learning has been the predominant paradigm by twenty twenty-one to twenty twenty-five. That’s where most of the current research is focused.

Jane: That leads nicely into their discussion on terminology and what exactly they mean when they say "intermodal loading unit identification." They’re proposing a standard way to talk about this task for everyone involved.

Paper summary: Tom: They suggest using the term intermodal loading unit identification as an inclusive name for determining the ISO six three four six codes of these units. Then, downstream tasks are broken down into things like object detection, instance segmentation, character recognition, and character recognition itself.

Lu: That breakdown helps clarify the complexity; it shows that identification isn't just one single step but a whole pipeline of computer vision tasks.

Meng: I’m interested in the data part because I know how much of this kind of work relies on having good training data, and that seems like a huge hurdle.

Tom: They point out that the quality and diversity of annotated data really influence how well deep learning models perform when they are actually deployed in a real port environment. The problem is that most datasets aren't publicly available, which makes it hard for new research to get started.

Jane: So, the lack of public benchmarks is a major constraint they’re flagging right away. What does that mean for someone trying to build a new model today?

Tom: It means the median dataset size is often too small to robustly train deep learning models on its own. Researchers end up having to adapt general purpose text detection and recognition models, which isn't always ideal.

Lu: They also mention that evaluation metrics cover a lot of ground: you’ve got end-to-end accuracy, precision, recall, F1 score, average precision or AP, mean average precision or mAP. They even look at speed metrics like frames per second or FPS.

Meng: Speed is critical when you're talking about real-time monitoring in a terminal. If a model runs too slow on the hardware we have available, it just doesn't work in production.

Tom: Exactly, and they mention that they look at speed metrics like frames per second as well. That’s because deployment has to be efficient enough to keep up with the flow of traffic in a port operation.

Jane: So, moving on from what we know about the past research, what kind of challenges are they suggesting we tackle next? Where does the field go from here?

Tom: They suggest a few key areas for future work. First, they want to address the new challenges that come with vehicle mounted cameras. Second, they suggest exploring synthetic data generation to get better training sets without relying only on real-world data.

Lu: And they also talk about taking those multi-stage methods and refining them into unified end-to-end models to reduce complexity in the processing pipeline. That’s a big idea for streamlining how the system works.

Paper summary: Meng: From my side, I think focusing on contextless text recognition is important because ISO six three four six compliant ID codes are inherently contextless, which means they don't have surrounding text that helps a language model guess what they are.

Tom: That’s a sharp point about the ISO six three four six compliance. And another direction they suggest is looking at image enhancement techniques, like noise reduction or using generative adversarial networks to fix those damaged or degraded ID codes we see in the real world.

Jane: So, to wrap up their suggestions for future work, it’s a mix of better data generation and smarter model architecture refinement. What's the big picture impact they want these things to have?

Tom: They suggest that tackling these challenges can lead to minimizing human errors, decreasing turnaround times at the port, lowering operational costs, and increasing resilience against disruptions. That’s the whole goal of pushing this research forward.

Lu: It really boils down to making the identification process more robust and less dependent on perfect real-world conditions for every single deployment.

Meng: I just worry about deploying a complex unified model that might be brittle when it encounters an ID code that looks slightly different from what it was trained on, which is something they hint at as a limitation.

Tom: That’s fair, and the paper does mention that the contextless nature of ISO six three four six compliant ID codes makes this approach less effective for language model based encoders. It's a constraint they acknowledge.

Jane: So, to conclude this scoping review on automatic intermodal loading unit identification using computer vision, what's the final word on what this research tells us about the field?

Tom: The main thing is that the shift from static image acquisition settings to dynamic ones introduces new challenges that have real potential for logistics and transportation. Addressing those challenges can help minimize human errors and speed up processes significantly.

Jane: It sounds like they’ve laid a solid foundation by mapping out the existing methods and clearly pointing toward the data gaps we all need to fill next.

Lu: The paper highlights that while CV offers a way forward, the lack of public benchmark datasets and standardized terminology is currently hindering further advancements in this area.

Meng: That lack of public benchmarks is what keeps me up at night; you can't really test if your model will work reliably on actual port traffic without someone else having run a test case.

Tom: So, the takeaway from this scoping review is that precise monitoring through dynamic camera setups has transformative potential for port operations if we can overcome those current data and standardization roadblocks.

Conclusion: Tom: So we're wrapping up this look at automatic intermodal loading unit identification using computer vision with this scoping review by those authors between nineteen ninety and twenty twenty five.

Jane: It seems like they’ve mapped out how much research has gone into this, showing the whole journey from older static systems to the deep learning methods we see today.

Lu: They really catalog all sixty-three empirical studies on this task, which helps show exactly where the field is focused and what's missing.

Meng: What does that cataloging actually tell us about what’s working versus what’s just academic theory?

Tom: It tells us that the methods have shifted dramatically over thirty years, moving from simple optical character recognition gates to vehicle mounted cameras for precise monitoring.

Jane: And they highlight a big gap in the data side, which is important because real port environments are tough to get labeled for training deep learning models.

Lalam: The review emphasizes that the lack of public benchmark datasets is a major barrier, meaning new research has to start from scratch almost every time.

Tom: They propose some standardized terminology, like calling it intermodal loading unit identification, to give everyone a common language for this task.

Jane: It’s about making sure that when researchers talk about object detection or character recognition, they are all talking about the same thing in port operations.

Lu: The future work they suggest focuses on synthetic data generation and refining those multi-stage methods into unified end-to-end models to simplify things for deployment.

Meng: I'm still concerned about deploying those unified models if they can’t handle every single variation of an ID code that pops up in the real world.

Tom: They also point out that ISO six three four six compliant ID codes are contextless, which makes them tricky for language model encoders to understand directly.

Jane: So, the big picture here is that moving toward dynamic cameras has huge potential for reducing human error and speeding up turnaround times in logistics.

Lu: Yes, addressing those data and standardization roadblocks is what will actually unlock those improvements in how ports operate.

Emre Gülsoylu, Alhassan Abdelhalim, Derya Kara Boztas, Ole Grasse, Carlos Jahn, Simone Frintrop, Janick Edinger

University of Hamburg Department of Informatics Computer Vision Group University of Hamburg Department of Informatics Distributed Operating Systems Group Hamburg University of Technology Institute of Maritime Logistics Hamburg University

cs.CV

Submitted: 2025-09-22

Updated: 2026-11-15

Comments: Submission to Transportation Research Interdisciplinary Perspectives

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 80/100

The gist: Automatic Intermodal Loading Unit Identification using Computer Vision: A Scoping Review maps Computer Vision (CV) methods for ILU identification, clarifies terminology…

Key concepts

Intermodal Loading Unit Identification (ILU)
This is the task of determining specific identification codes for containers or loading units within a port environment. It involves various CV steps like object detection and character recognition to accurately read these codes from images.
Methodological Evolution
The techniques used to identify ILUs have changed significantly over time. Early methods relied on static systems, but the field has recently transitioned to more precise vehicle-mounted camera setups, with Deep Learning now being the dominant approach for identification.
Data and Evaluation Challenges
A major hurdle in this research is the scarcity of high-quality, publicly available annotated data. Small dataset sizes limit the training of robust Deep Learning models, forcing researchers to adapt general models. Evaluation involves metrics like accuracy and speed (FPS).
Recognition in CV
In computer vision, recognition specifically refers to extracting text from an image. For ILU identification, this means accurately reading the specific alphanumeric ID codes present on a container or loading unit using techniques like OCR.

Terminology

Summary

Automatic Intermodal Loading Unit Identification using Computer Vision: A Scoping Review maps Computer Vision (CV) methods for ILU identification, clarifies terminology, and highlights research gaps to inform future work in port operations

Scope of the Review

The review followed PRISMA-ScR guidelines to map empirical studies on CV-based solutions for ILU identification published between 1990 and 2025 The search yielded 63 empirical studies on CV-based solutions for the ILU identification task, published between 1990 and 2025 The review catalogues all empirical studies on ILU identification using CV methods between 1990 and 2025

Methodological Evolution

The methodological evolution of ILU identification solutions shows a shift from static (e.g. OCR-gates) to vehicle mounted camera setups, which enables precise monitoring The reported results for end-to-end accuracy range from 5 % to 96 % The development of methods for ILU identification has undergone a significant transformation over three decades, as illustrated in Figure 4 For two decades between 1990 and 2010, the field relied predominantly on DIP techniques combined with early ML methods and to a lesser degree, DIP methods alone The five-year interval from 2016 to 2020 marks a turning point, with a notable adaptation of DL-based methods By 2021–2025, DL-based approaches had established themselves as the predominant paradigm for ILU identification

Terminology and Tasks

The review proposes standardised terminology to create a common ground for researchers with different backgrounds The term Intermodal Loading Unit Identification is proposed as an inclusive designation for the task of determining ID codes of ILUs Downstream tasks include object detection, instance segmentation, character segmentation, character recognition, and OCR The term recognition refers to extraction of text from an image in the CV field

Data and Evaluation Challenges

The quality and diversity of annotated data strongly influence how well DL models perform once deployed in real-world port environments The significant majority of the datasets are not publicly available, which creates a substantial barrier for research and development in this area The median dataset size is too small for robust DL training, forcing researchers to adapt general purpose text detection and recognition models Evaluation metrics include end-to-end accuracy, precision, recall, F1-score, average precision (AP), mean average precision (mAP), character error rate (CER), and speed/efficiency metrics like frames per second (FPS)

Future Research Directions

Future work suggests addressing new challenges emerged from vehicle mounted cameras, exploring synthetic data generation, refining the multi-stage methods into unified end-to-end models to reduce complexity Researchers should focus on contextless text recognition due to the nature of ISO6346-compliant ID codes Other directions include multi-view ILU identification, image enhancement techniques such as noise reduction or GANs to address degraded or damaged ID codes Refining network architectures to increase the receptive field and incorporating attention mechanisms are also suggested for DL-based methods The development of unified, end-to-end models is suggested to streamline processing pipelines

Conclusion

The shift from static to dynamic image acquisition settings introduces new challenges that have transformative potential for transportation and logistics Addressing these challenges can result in minimised human errors, decreased turnaround times, reduced operational costs and higher resilience against disruptions Precise monitoring can optimise the terminal operations, reduce human errors, pave the way for faster turnaround times, lower operational costs and increased resilience against disruptions like misplaces ILUs The shift from static to dynamic camera settings introduces new challenges that have transformative potential for transportation and logistics The contextless nature of ISO6346-compliant ID codes makes this approach ineffective for language model based encoders

Keywords

intermodal loading unit identification, container code recognition, lack of public benchmark datasets, open-access code, and standardised terminology hinders the advancements in this field The shift from static to dynamic camera settings introduces new challenges that have transformative potential for transportation and logistics The lack of public benchmark datasets, open-access code, and standardised terminology hinders the advancements in this field The shift from static to dynamic camera settings introduces new challenges that have transformative potential for transportation and logistics The lack of public benchmark datasets, open-access code, and standardised terminology hinders the advancements in this field The shift from static to dynamic camera settings introduces new challenges that have transformative potential for transportation and logistics

References

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Standardize terminology to create a common ground for researchers across disciplines, which will help in meaningful cross-study comparisons by clarifying terms like text spotting versus OCR.

  2. Implement end-to-end models that unify detection and recognition stages to reduce complexity, potentially leading to minimised human errors, decreased turnaround times, reduced operational costs and higher resilience against disruptions.

  3. Develop contextless scene-text recognition models specifically for ISO6346-compliant ID codes, as current semantic context relying on natural language models is not the most suitable approach for these codes.

Abstract

Background: The standardisation of Intermodal Loading Units (ILUs), including containers, semi-trailers, and swap bodies, has transformed global trade, yet efficient and robust identification remains an operational bottleneck in ports and terminals. Objective: To map Computer Vision (CV) methods for ILU identification, clarify terminology, summarise the evolution of proposed approaches, and highlight research gaps, future directions and their potential effects on terminal operations. Methods: Following PRISMA-ScR, we searched Google Scholar, dblp and OpenAlex for English-language studies with quantitative results and charted the included studies across methods, datasets, failure modes, and evaluation metrics. Results: We reviewed 95 studies published between 1990 and 2026. A shift from static (e.g., OCR-gates) to vehicle-mounted camera setups is observed. Only 26 studies report an end-to-end accuracy (median 94.5%, range 19-100%). Eight datasets are publicly available, yet none is commonly used as a benchmark, and six studies describe a deployed system and seven a field test. Conclusions: We propose the term ILU identification for this broader task, which includes finding the unit and its code and covers semi-trailers and swap bodies. The literature still reflects its origin at the gate: challenges typical of moving cameras receive the least attention, and no study reports separate results for semi-trailers or swap bodies. Evaluation requires public benchmarks that cover these conditions and all ILU types, with end-to-end accuracy results. Promising directions are synthetic data, unified end-to-end models, recognisers built for codes without linguistic context, and mission planning with automatic coverage path planning for vehicle-mounted cameras. Addressing these challenges could reduce human errors, turnaround times and operational costs and increase resilience against disruptions.

Sources

Related papers