The team behind our AI and deep learning fueled engines
AN INTERVIEW WITH CAMCO’S IMAGE ANALYSIS TEAM WHO ARE CURRENTLY APPLYING AI AND DEEP LEARNING TECHNIQUES TO TACKLE A VARIETY OF PROBLEMS
A special unit within Camco’s software engineering department is the Image Analysis team, which counts six computer vision experts, highly skilled in artificial intelligence (AI) and deep neural network fundamentals. It’s no surprise that this geek club speaks its own lingo and keeps to Camco’s basement. Let’s find out what they’re currently working on and what makes them tick. An interview with Peter, Milan, Ottokar, Klaas, Joris and Hendrik.
What is the most important prerequisite for visionbased AI applications?
Peter: With twenty years of experience in this domain I can say that it all starts with the quality of an image, meaning detail, sharpness, varying angle, sun/shadow and day/night effects.
Whatever technique you use, old school image analysis or deep learning, a bad picture gives poor results. So, I can say that I’m a happy developer, since Camco puts a lot of effort and investment into making the best pictures in the market, in sometimes difficult outdoor environments. Since we develop our own cameras, this gives us more flexibility in improving the image quality.
Can you tell us more on how Camco evolved from traditional image analysis techniques to deep learning methods?
Klaas: Since Camco’s early years (2000s), there’s been lots of progress in image recognition and we’ve always tried to stay on top of things. Our first OCR engines for example were based on pixel (intensity value) comparison techniques such as template matching and principal component analysis (PCA).
This proved to be a satisfactory approach for recognizing characters, but not for more complex matters such as danger-
ous goods labels and container seals.
Later methods involved scanning the images for distinct patterns or features that characterize an object, such as its shape, color or texture, and classifying the image based on those features. One of these methods, HOG or Histogram of Oriented
Gradients, started to be widely used for object detection: in other industries for detecting pedestrians and traffic signs, here at Camco for seal and IMO detection.
In the meantime, research in deep convolutional neural networks (CNN) was steadily growing and deep learning completely took off in the 2010s. Yet for industry purposes there were still some remaining difficulties regarding the required computing power, scalability and big datasets. About four years ago, Camco was able to start using deep learning methods to solve a variety of machine learning tasks and has been booking excellent results ever since.
How does this work exactly?
Klaas: Convolutional neural networks learn by experience. This means that instead of handcrafting patterns or features to look
for in images, we simply feed the network a large number of labeled images and it will self-learn the features that characterize an object.
It’s not something you can easily visualize, because what has actually been learned are just values of millions of weights and biases that form the connections of a multi-layered network. Feature extraction in each consecutive layer based on these
learned weights ultimately results in a practical output: the object’s class and its location(s).
Camco is successfully using deep learning for better OCR recognition rates and a lower number of exception jobs.
Our camera software currently run engines for:
- (streaming) License Plate OCR
- Chassis OCR
- Wagon OCR
- Seal & IMO detection
- Container & trailer number OCR
- Container / cargo detection & classification
- Wheel detection for axle counting, waste label detection, … Refer to Camco Times 6 issue for more details.
Ottokar: This process of supervised learning requires a large training set and coming up with the best suited network architecture (i.e. a definition of layers and operations) can be very time-consuming. However, very rarely will we build and train our own models from scratch. As numerous research groups frequently release state-of-the-art algorithms and CNN architectures for a wide range of problems, we are happy to pick the fruits of their published work and pre-trained models, and
tweak the network architectures in our advantage.
At Camco, we closely follow up on trends in our industry in order to stay on top of new evolutions and steer our development in the right direction.
So, where’s Camco’s added value and expertise?
Milan: Our strength lies in analyzing the circumstances for each problem, gathering high quality datasets and optimizing (existing) models and training procedures given time and memory constraints. Not to mention a great deal of hand-coding to get the job done and meet specific customer requirements. This is an iterative process. We are continuously improving our methods and models as data keeps flowing in and new discoveries are made in the field of deep learning.
Hendrik: Let’s not forget that we get support from Camco’s Data Analysis team who make the statistics—they check our
algorithms in daily use.
Can all image recognition problems be solved with deep learning nowadays?
Milan: Deep learning is definitely driving today’s AI explosion and is considered its most powerful tool. But it’s not the only tool in our shed because deep learning alone is not always sufficient to tackle problems. Sometimes current resources and costs, lack of samples, or the complex and unpredictable nature of a problem are limiting factors and a combination of technologies and algorithms might be required. That’s when our twenty years of experience really pay off.
Camco’s customers are typically terminal operators, facing new challenges as their business is continuously developing. Can you tell us more about typical customer requests that come your way?
Milan: A first example is related to seal types. Although our software was able to detect bolt seals on container and trailer
doors, a specific customer wanted to know the seal type. By expanding our datasets and retraining our CNN-classifiers, we
are now able to determine the presence of other seal types as well.
Ottokar: An intermodal terminal operator asked us if we could measure the thickness of a railcar’s brake shoes, to automatically detect when they’re worn down and due for replacement.
The main challenge here is the actual image capturing of all of the brake shoes, as they’re—often—hidden by the wagon body or other hardware.
Another request was to perform real-time tracking of dock workers at work under the crane to enhance terminal safety. By fusing data from multiple sensors, namely vision-based AI for person tracking and our proprietary micro location systems
for asset tracking (e.g. straddle carriers), we can flag dangerous situations and generate alerts.
Milan: A recurring customer request is to perform automated damage inspection (ADI) on containers and trailers. As traffic
continuously increases, so does the number of accidents and damages. Obviously, manual damage inspection is very timeconsuming and impractical for large terminals.
For one of our customers, a multinational logistics company, our team has been busy developing solutions for vision-based ADI, where the focus lies on detecting damage seen on both the container/trailer body, as well as on the chassis. Below are some examples of trailer damage next to their corresponding ADI heat map.
What about laser technology for ADI?
Peter: Unfortunately, a camera cannot see the difference between a black spot and a hole, for example on a container’s surface. Laser scanning could be useful as it performs depth scanning on a surface. However, we’ve concluded that at the moment, laser technology is not yet capable of examining container or trailer surfaces at an acceptable speed and resolution. As current lasers have a maximum scan speed of 100 Hz, this means that we can scan a truck driving at 10 m/sec (36 km/hr, 22.4 mph) only once every 10 cm, leaving most of the object unchecked. As a result we cannot reliably detect damages with a size under 10 cm.
What’s more, the laser can only detect deformations of the object’s surface—scratches and small dents remain undetected. Also, the laser signal depends on the reflectivity of the object and we found out that very dark objects often won’t return enough signal for damage inspection. And finally, there’s the price tag. Lasers are very expensive and we would need at least one laser for every side of the truck (left, right and top). It’s also very difficult if not impossible to scan the front and the back of the truck or trailer while the truck drives through the gate.
We concluded that since we already use cameras to capture the images, it makes sense to use them for damage inspection as well, saving the need for extra hardware.
Ottokar: Since we have opted for a complete image-based solution, our first and foremost requirement is capturing the damage, clearly visible to the human eye, regardless of weather and lighting conditions. This challenge encourages Camco to continue innovating its lighting and camera systems.
Next, we need to think of combining existing solutions and new developments. ADI is a very comprehensive concept and is to be divided into dedicated components for each subproblem. There is no almighty deep learning model that processes everything. For example, if we want to check if a truck’s mud flap is missing, we first need to identify a region of interest. This
is done by detecting the wheels and then processing the adjacent areas.
The same holds true for body damage. It’s good to know which area of the trailer or container we’re looking at in order to call
the designated engine and rule out certain outcomes.
AI and deep learning in particular require a lot of computing power. How do you run these applications?
Joris: Camco advocates local processing, since moving images across networks is very time consuming and creates many
dependencies with respect to timing and reliability. All the latest generation Camco intelligent cameras have fast Intel i3 processors embedded, which offer more than adequate performance (within seconds) for most of our deep learning applications.
Nevertheless we do have applications, such as streaming LPR, that require faster processing. So we integrated a Jetson TX2
NVIDIA module in some cameras. This is specialized hardware designed for AI computing. With this module, we are able to run our software on a GPU. This allows us to use deeper and more accurate neural networks on cameras, and still meet our
time constraints, thereby increasing our recognition accuracy.
Benchmarking showed that running our current (lightweight) networks on a GPU results in only a small improvement in
computation time—partially because current code is already heavily optimized. However, the performance increase is much
more noticeable when using deeper networks. The challengelies in approaching problems from a different angle to unlock
the full potential of GPU accelerated computing.
Looking at the future
Camco is facing increasingly tougher challenges in image recognition that require new and advanced solutions. The use of
higher capacity neural networks, which were previously out of the question due to time and memory constraints, now lie
within our reach, thanks to our hardware innovations.