INLOC Robotics

Computer Vision for detection and (multi) tracking of circles in a video

The tracking of objects in a video is a common goal in computer vision. Some methods rely on the search of keypoints, which are then matched with keypoints on the next frames.

In this post we will explain a method based on the identification of simple shaped objects, in particular circles, to track them in a robust manner (i.e. despite the movement or change of size) for all the frames in video where the object appears. In addition, the method allows the tracking of more than one circle simultaneously (multi-tracking).

The case of study is a static scene and a moving camera, but a similar method could be designed for static camera and moving objects. This method is divided in two parts: first the identification of circles and then the tracking.

Circle detection by computer vision

The circles are searched frame by frame, without any knowledge about circles detected in previous images. There can be more than one circular object in the same frame, so the circle detector must be able to return several circles, if that is the case. Therefore, each frame provides a set of candidates of being a circular object in the image.

The frame is first processed so that circular shapes can be easily found. The most important step in pre-processing is to obtain the borders of the objects. From those borders, circles are searched and only the best are returned.

At this point, some circles could not belong to circular objects, so a post-processing is also needed. Circles may be filtered according to our previous knowledge about the objects, for example by setting a bound on accepted radius or positions. We can also check if circles are crossing between them, or whether they are too close or too far one from another.

Computer Vision Based Circle Detection - Gray

Computer Vision Based Circle Detection - Sobel

Grayscale image, edges and circles detected (red). The blue and green circles are the tracks.

For this part, as important is to detect the circles when they appear in image as not detecting when there are no circular objects. Therefore, only the best circles, if any good, are returned for the tracking part.

(Multi) tracking of circles

The goal of multi-tracking is to estimate the position (center) and size (radius) of circles along all the frames where the object is visible, simultaneously for all the objects. This estimation must be computed even if there is no circle detected for one object in one frame (or several consecutive frames).

For those cases when a detection is missing, a dynamical model has been developed for prediction of size and position of circles in consecutive frames, by knowing the movement of the camera (in the case of static scenes).

As long as new circles are detected as object candidates, they are either discarded, linked to a current tracked object or starting their own new track. The tracking has been shown to be robust against wrong circle detections.

Circles are discarded if they are not similar or different enough from tracked objects. A circle is set as a new observation of one object if it is similar enough to the prediction for that object.

Finally, a new object is detected if the detected circle is different enough from any predicted circle for all the current tracks.

After some of the tracked objects have been linked to a detected circle, the state of tracked objects is estimated. This is done frame by frame, combining the information provided by the prediction model and the circle observed, resulting in a smooth and complete knowledge for the object position and size along all the frames.

Conclusions

This method provides an accurate estimation of several circular objects simultaneously, computing an measurement of size and position for all the frames, even if circles are not always detected. It is robust against the change of size and position of the objects, and can be extended to similar methods involving other shapes and scenes.

We hope this article has been useful to you. If you have an engineering project in your hands and you think we can help you, here is the link where you can contact us and explain more about it.

Contact

Industrial use of GPUs

GPUs (Graphics Processing Units) were born due to the need to quickly manage the interface currently displayed on computers. As time goes by its use was extended to graphic design programs, video games, image processing,… But in this article we will focus on the capacity for numerical processing and artificial intelligence.

GPUs could be understood as a large number of small units capable of doing simple calculations simultaneously. This, although it seems simple, gives them exceptional processing power, it is not the same to do 1000 multiplications one after another than to do them all at once.

This capacity allowed the development of the first AIs and, in general, to introduce GPUs in research projects. Some examples would be the complex fluids simulations, calculations related to quantum chemistry, paralyzing calculations on large amounts of data…

Today, more and more companies choose to use GPUs in their systems, without going too far, facial recognition applications that detect faces on our smartphones use their small GPU, or the self-labeling of Facebook images would be two examples.

At INLOC Robotics we are not far behind in this field, we have our own AI within the SewDef system and we implement parts of our algorithms on GPUs to optimize them to a great extent. This opens the door to the increasingly frequent use of this hardware in industry.

As a practical example, we can use the case of IoT. In many companies the use of this system is being implemented to monitor production in real time.

Now let’s consider, once the first IoT systems are implemented, these can be increased exponentially by adding more data to the system. More computing power is likely to be required, which in turn allows for more elaborate studies of the production. The way forward would be the use of GPUs, in order to maintain the real-time monitoring that characterizes the IoT.

Breaking weaknesses:

When it comes to GPUs for general use, we can see two major weaknesses. First of all, GPUs were not initially designed with this use in mind, causing a program designed for a GPU to not work with the same performance on the next generation of the same hardware.

Another recurring problem is the lack of precision in calculations with decimals. To increase speeds, the precision in the same hardware is usually reduced, this fact is not a problem for all applications, but it could be.

However, the growing demand from the research sector of GPUs has made manufacturers focus efforts on solving these weaknesses, to the point of reducing them enough so that they are already used today in a wide variety of devices and/or applications.

In my opinion, we have reached a turning point where GPUs will have a strong entry in the industrial sector solving very diverse problems.

We hope this article has been useful to you. If you have an engineering project in your hands and you think we can help you, here is the link where you can contact us and explain more about it.

Contact

INLOC Robotics CTO Presentation on World Water Day 2021

Within the framework of the celebration of International Water Day on Monday March 22, 2021, the CTO of INLOC, Josep M. Mirats Tur, was invited by the Canal Foundation to give a talk on the application of Artificial Intelligence for thedetection of sewer defects from CCTV camera inspection images.

In this talk we explain the main advantages of the use of AI for the detection of defects, allowing the obtaining of objective information that will allow infrastructure managers a better decision-making regarding investments dedicated to sewer maintenance.

If you want to know more about the event, you can go to the event page by clickinghere.

We hope this presentation has been interesting to you.If you want to know more about SEWDEF and the automation of detection of defects in the sewer system, here is the link where you can contact us and find out more about all the benefits that you will obtain with SEWDEF.

Contact

Buried Pipe Inspection and Sewerage Management

How much does inspecting buried infrastructure cost? And perhaps more importantly, how much does it cost not to do those inspections? Can Artificial Intelligence help us with sewer inspection?

We live in a world where urban agglomerations have become the majority ecosystem of the human population. More and more people live in urban environments, and cities are getting bigger and bigger. And for these ecosystems to be sustainable for human life, they need to be equipped with infrastructures that allow access to such basic needs as safe water and sanitation. And these infrastructures are becoming more complex, difficult to design and build, and challenging to maintain.

Resources for Infrastructure Management

Resources for infrastructure management and maintenance are always finite, and it is imperative to make the right decisions when prioritizing where to apply these resources to avoid damage and accidents with consequences that can be catastrophic.

Over the past few years, improved sensor technology, data acquisition, and IT technology has led to the birth of better infrastructure management systems with the gradual incorporation of Asset Management and the paradigm shift from corrective inspection to preventive inspection.

However, buried infrastructure such as sanitation and sewerage infrastructures still requires regular robot camera inspections to check the actual condition of the infrastructure. After all, management systems will give information as good as the data they are fed with. And the amount of data to be analyzed can become a difficult bottleneck to avoid.

SEWDEF and Automatic Sewer Defect Detection

An automatic sewer defect detection system, through analysis of the images obtained in the inspection, becomes an artificial assistant that makes it possible for this preventive inspection of thousands of miles of buried sewerage to be now possible and feasible.

Thanks to an artificial intelligence system optimized for sanitation pipes, large volumes of data can be analyzed for objective, reliable, time-consistent, and standardized results.

Operating costs are significantly improved, and appropriate decisions can be made in applying resources where they are most needed in order to maintain the operation of the infrastructure at an optimal level and avoid direct and indirect costs arising from stateless and accelerated deterioration.

SEWDEFis an automatic sewer defect detection system that, through a cloud-hosted artificial intelligence system, robotic self-localization algorithms and “Deep Machine Learning” enables the automatic identification and localization of defects in sewer pipes quickly and effectively. With SEWDEF, infrastructure managers can now integrate real state data into their predictive algorithms and radically improve their decisions when using their limited resources.

We hope this article has been useful to you.If you want to know more about SEWDEF and the automation of detection of defects in the sewer system, here is the link where you can contact us and find out more about all the benefits that you will obtain with SEWDEF.

Contact

Evolution of Artificial Intelligence

We live in a time when technology advances by leaps and bounds and in the middle of it we find Artificial Intelligence, something only theoretical in the past, expanded in almost all aspects of our life in the present and a hope for the future.

What is Artificial Intelligence?

In the past, Artificial Intelligence has been described as: Any task performed by a machine that was previously considered to require human intelligence.

In the current we got new definitions such as: the ability of a system to adapt and improvise in a new environment, to generalize its knowledge and apply it to unknown scenarios.

Timeline of the Artificial Intelligence

During the more than 70 years of AI history, there have been 3 great eras separated by the central theme of their study and development:

1950-1970: Development of AI algorithms and methods.
1970-1990: Development of symbolic algorithms and expert systems also known as knowledge systems.
1990-Present: Machine learning and deep learning.

From these methods we can also divide between Model-Driven development and Data-Driven development:

Classical programming vs machine learning

Where the MD need a series of instructions to arrive at the solution (for example using decision trees).
And the DD use data with their solutions to “train” the system and the system provides the necessary rules so that when we input new data it answers with the corresponding solutions.

Uses of Artificial Intelligence

Artificial intelligence tries to imitate human behavior and for this it uses:

Robotics: Helps control systems that allows machines to manage physical tasks such as movement while adjusting to changes in the environment.
Natural Language Processing: Some applications include information retrieval, text mining, answering questions, and machine translation.
Computer Vision: can be used for object recognition, event detection, scene reconstruction (mapping) and image restoration.

With the above fields AI is used today in many sectors such as:

Virtual Assistants (Siri, Alexa, Cortana …)
Finance (Trade and investment, auditing, underwriting …)
Health (Scanning of images, sound analysis of the heart …)
Service sector (HR and hiring, job search, marketing)
Media and e-commerce (Deepfakes, music, videogames …)
Transportation (simulators, autonomous vehicles, route planning …)
Industry (damage analysis, sediment monitoring …)

Future of AI

AI is already integrated into our society and it is apparent that it is not going away anytime soon. Which ends up leading us to the next question: What does the future of AI hold for us?

Transparent AI:

Deep learning is made to replicate in a similar way neural networks and, like our brains, they receive data input and return an output but we do not know how they arrive at that solution. By not knowing how neural networks identify patterns and derive results, it is difficult to replicate and improve those results.</p

For this reason, the creation of a Transparent AI in which we can identify how our system reaches these solutions can suppose an exponential evolution for the development of AI.

Strong Artificial Intelligence:

All the AIs that have been created so far are part of the Weak Artificial Intelligence or Narrow AI type and work by focusing and solving specific tasks.

Instead the SAI (also called Artificial General Intelligence) is the AI that equals or exceeds the average human intelligence. This type of AI has the ability to use reason, represent knowledge, plan, learn, communicate, and integrate all these capabilities to solve a problem.

At the moment this type of AI is still far from being achieved but advances in technology and science continue to evolve exponentially and what years ago seemed like science fiction is now a reality.

Digital fingerprint reader

A fingerprint reader was until recently quite an exotic technology in the real world, where we watched futuristic movies with these types of devices.In recent years, however, scanners have started to appear everywhere: high-security buildings, mobile phones, office entrances, and even on PC keyboards.

Fingerprints are a unique marker for each person.While two prints may look basically the same, advanced software can detect sharp and clear differences.

In this article, we will talk more about the fingerprint detector and the different types out there.

What is a fingerprint reader?

Fingerprint sensors are electronic devices that allow us to identify our fingerprints.This process is mainly composed of reading, saving and identifying the fingerprint.Each imprint of our fingers are different from each other, and of course, different from anyone else.

How does the fingerprint reader work?

A fingerprint scanner system has two basic functions.On the one hand, you need to obtain an image of your finger and on the other, you must determine if the pattern of ridges and valleys in that image matches the pattern of ridges and valleys in previously scanned and saved images.

Fingerprint scanning is a form of physiological biometrics that analyzes your physical characteristics to authenticate your identity.Basically, it recognizes that your fingerprint belongs to you and no one else.

We all have unique identification marks on our fingers that are used to create a fingerprint.These cannot be changed or deleted, so they are a good indicator of identity for security procedures.

Keep in mind that some scanners rely on light, some on electricity, and some on sound to map the ridges and valleys of your fingers.

Types of fingerprint scanners

There are four types of fingerprint scanner:

Optical scanner
Capacitance scanner
Ultrasonic scanner
Thermal scanner

Optical scanners take a visual image of the fingerprint using a digital camera. The set of pixels form an image of the scanned scene (in this case, a finger). Subsequently, the captured fingerprint is compared with the registered fingerprints.

Capacitive or CMOS scanners use capacitors and thus form an image of the fingerprint through electrical current. This type of scanner tends to stand out in terms of accuracy. These types of scanners are usually more compact than optical devices.

Ultrasound fingerprint scanners use high-frequency sound waves to penetrate the epidermal (outer) layer of the skin. The waves are then reflected off the sensor which are then analyzed to create a digital image of the fingerprint.

Finally, thermal scanners detect temperature differences at the contact surface, between the ridges and valleys of fingerprints.

Advantages and disadvantages of a fingerprint reader

Fingerprint scanners have a number of advantages over other systems.Below we name a series of them.

Physical attributes are much more difficult to counterfeit than identity cards.
You can’t guess a fingerprint pattern like you can guess a password.
You can’t lose your fingerprints, iris, or voice like you can lose an access card.
You can’t forget your fingerprints like you can forget a password.

But as effective as it is, certainly a fingerprint sensor is not foolproof and they have major disadvantages.Optical scanners cannot always distinguish between the image of a finger and the finger itself, and capacitive scanners can sometimes be fooled by a cast of a person’s finger.

Even in the worst case, a criminal could even cut off someone’s finger to get past the scanner’s security system.

To make these security systems more reliable, a good idea is to combine biometric analysis with a conventional means of identification, such as a password (in the same way that an ATM requires a bank card and a PIN code).

Conclusions

There are several ways that a security system can verify that someone is an authorized user.

In general, to overcome a system you look for “what you have“, such as an identity card with a magnetic stripe.On the other hand, it looks for “what you know” and requires you to enter a password or a PIN number.Or finally, “who are you“, where you look for physical evidence that really confirms your identity and it is here where we find a voice pattern, an iris or specifically the main topic of this article, a fingerprint.

We hope this article has been useful to you.If you have an engineering project in your hands and you think we can help you, here is the link where you can contact us and explain more about it.

Contact

Segmentation and decoding of text in an image

In the world of computer vision, it is very common to find images where, among other objects or environments, text also appears. Sometimes, it is of special interest to be able to read it, so its segmentation from the rest of the image is of great importance.

The degree of difficulty of text detection varies greatly depending on the environment. That is, it is not the same to detect text in a “controlled environment” where the text position is known and it is clearly differentiated from the rest of the image, than in a “natural environment”. In the latter, a series of factors interfere that greatly hinder the segmentation of the text, such as the noise of the camera with which the image is obtained, the poor lighting of the scene or the blurred frames that occur, for instance, if the camera is not stable.

In addition to the problems already mentioned, there is also the difficulty of locating the text within the image since it can appear in different positions and orientations. Once it has been located, each character must be carefully segmented in order to obtain a correct reading of the text.

Detection of text in an image

As mentioned above, the first challenge to face in text segmentation is the location of the text. Among the different possible methods to achieve this goal, in this case the EAST Detector will be used.

The EAST Detector is capable of detecting text practically in real time (13 fps) in both images and videos, whether in horizontal or rotated text using a convolutional neural network.

With respect to other possible algorithms, the EAST Detector has eliminated unnecessary intermediate steps so that it only has two steps. The first one is the prediction of lines of text or words using the neural network and the second one is the processing of predictions.

In the upper Figure you can see an example of the different regions with text that have been detected by the algorithm, each one marked with a green bounding box.

Decoding the text

Once the location of the text has been detected, it has to be decoded. For this, it is important to be able to isolate in the best possible way the characters from the background of the image. Therefore, different morphological operations must be applied to achieve this. These operations depend on the environment with which you are working, so the best is to evaluate each case individually to decide which ones should be applied. The Figure below presents an example of this step.

Once the text has been isolated from the background of the image, there are different methods to be able to read it. In this case, we have chosen to use the Tesseract OCR library, which is an engine for optical character recognition.

Thus, the combination of the EAST Detector together with the Tesseract library provides a fairly robust method by which the text position can be detected and read for later processing of these data.

There are different options that allow you to further refine your character recognition. For example, it is possible to indicate the language in which the text is found or if it is alphanumeric characters or exclusive numbers or letters.

Project ASIR collaboration: Sewer robots autonomous navigation

INLOC Robotics SL has been part of the Danish ASIR (Autonomous Sewer Inspection Robot) project for two years. The main objective of this project is the design of an autonomous robot, capable of navigating through the sewer and automatically detecting the defects that can be found.

Recently, during the months of November and October, a close collaboration was established with the Danish University of Aalborg (AAU). The objective was the acquisition of real data from a sewer system built specifically for this project, using a prototype robot capable of capturing 3D data.

Afterwards, by INLOC Robotics SL, the acquired data will be used for the design of localization, obstacle avoidance and sewer structural elements detection algorithms. As for the AAU, the data will be used for the final robot sensors selection.

Sewer data collection using 3D sensors:

Data collection was carried out by Ferran Plana (INLOC Robotics) and Chris H. Bahnsen (AAU) in Aarhus, Denmark. The experiments were divided into 3 parts, localization, avoidance/detection and illumination.

For the localization experiment, a series of evenly spaced perforations were made in a tube by adding cylinders into them. In such a way, while the prototype robot captures images, the cylinders will help us to create a grown truth.

The grown truth will be used to verify the effectiveness of the different localization algorithms that the final automatic sewer inspection robot will use.

To obtain data for the obstacle avoidance algorithm, typical elements were used within the sewer, roots, rocks, sand… Taking different scenes with these elements, it is expected to be able to detect these in a 3D environment, which would allow to compute the evasion strategy for the future algorithm.

Regarding the structural elements detection, a similar strategy was followed. The prototype robot was placed in positions where there were lateral connections, different types of cameras, joints… The 3D data acquisition was varied enough to prepare 3D detection algorithms.

Finally, the prototype robot has two 3D sensors of different technology, a PicoFlex (time of flight) and a RealSense (stereo image). The RealSense camera requires lighting in order to take 3D images. The question would be: how much?

Using an industrial adjustable torch, the robot was positioned in the sewer within the dark. Modifying the illumination power at fixed intervals, different images were taken in each one of them. This will make it possible to create an illumination model, which will help to choose the minimum illumination necessary for a stereo-type camera, while maintaining its precision to the maximum.

Everything is ready for the development of the first autonomous navigation versions of the final ASIR project robot. Stay tuned for updates from INLOC Robotics SL to follow the evolution of this great project! An automatic inspection robot is a highly novel product that would greatly improve the efficiency and quality of sewer system maintenance.

Object detection and inference using a point cloud obtained with a 3D camera

The detection and identification of objects and people is one of the major points of contention in the implementation of systems with a certain level of autonomy, and it is undoubtedly one of the issues with the greatest impact in the world of computer vision.

Correct detection and inference is key to be able to carry out tasks such as autonomous navigation, robot arm assistance, video surveillance or, as in our case, defect detection.

There are, fundamentally, two ways to approach object detection.

The first is based on the use of monocular cameras (normal cameras, like those of our mobiles), which provide two-dimensional images.A series of algorithms are applied to these images, which can be very varied, in order to obtain a “box” or delimitation that indicates the presence and location of the object to be detected in the image.

The second is based on the use of 3D cameras that, as the name suggests, in addition to the 2-dimensional image, add a third dimension: depth.Thanks to this we can create point clouds, like the one in Image 1. In INLOC we are using this second method to perceive the environment.

The algorithm applied to the point cloud

The algorithms that can be applied are very varied, from the simplest to the most complex.

In this post we want to show the result of detecting a cup of coffee by concatenating two very simple algorithms: first, we apply an edge detector to the original point cloud.That allows us to expose the basic geometry of most of the elements;and second, we pass it through a RANSAC algorithm, which will make an estimate of the parameters of the geometry that best adapts to the mentioned edges of the point cloud.

Image 1. The point cloud as obtained from the 3D camera. Note the cup of coffee on the table, with the wall and objects in the background as noise

Image 2. The original point cloud with the detected and located coffee cup. The original dots are in blue, the edges are pink, and the dots that fall inside the box (in red) that encompass the mug are in gray.

Which geometry to use will depend on the problem.

In the case of Image 1, the objective is to detect the cup of coffee on the table, so estimating the parameters of a circle is the most logical thing to do.

By using this method, a successful detection of the cup is achieved, as can be seen in Image 2. The good thing about this method is that, if the object to be detected is reasonably simple, the detection is not only simple but also fast.

It is not, however, robust: when there are objects with complex geometries or many instances of similarly shaped objects in the point cloud, the chances of finding unwanted objects are high.

For these cases, hybrid solutions must be explored, in which the RGB image of each candidate object is passed on to other algorithms such as convolutional neural networks, which have greater capacity for abstraction and generalization, despite being much more complex and slow.

But that’s a story for another post!

Gestión de activos eficientes para el sector del agua

Efficient asset management for the water sector

The main objective of the digital transformation of an industrial sector is to provide tools that allow the total connection between its elements so that it can act in an agile way and, even more importantly, focused on the client.This process requires a great internal effort to manage change since it completely transforms the way of working and, consequently, the mindset of all its staff.

The water sector, probably due to the high implications of responsibility regarding citizen service, is relatively conservative in this process of change, as complex as it is necessary, in the way it works and manages assets.As a consequence, it presents a low degree of digitization compared to other productive sectors.

There is a clear tendency for users to demand more information in real time, they are less tolerant of service failures and the resources available to management companies are finite and limited.Thus, it is necessary to implement monitoring systems that allow the user to obtain data -especially from a good such as water-, on the one hand, and on the other, the managers have information to carry out maintenance taskspreventive and, above all, plan it properly taking into account the needs of the network and the resources available.

Examples of digital transformation in the water sector

Although it is true that in recent years the use of technology in the management of infrastructures in the world of water has become popular, coinciding with the explosion of Smart Cities, we observe that most of the digitized processes belong to the management of thedrinking water network.We have as examples the use of digital twins, the mapping of infrastructures through GIS systems, the use of big data to offer value-added services to the user, automatic sampling points for the determination of quality … This is due, as I mentioned.previously, to the demands of the user, who is increasingly demanding of information and is increasingly aware of what he consumes.

The digital transformation has also reached the Wastewater Treatment Plants (WWTP), which have highly sensorized and digitized control panels and processes to comply with the discharge regulations, as well as meet environmental objectives through energy optimization.

However, in this process of transformation and modernization of the sector, started in the visible parts of the integral water cycle, it seems that we have forgotten about the buried infrastructure, the sewage system.This has led to the continuing occurrence of overflows or spills into the sea in times of heavy rain.

It is here where our product SEWDEF, an automatic sewer inspection analysis system, offers objective information for efficient sewage management.Inspections of the sewerage network are carried out using images recorded from TV cameras that are manually analyzed, giving rise to a report subject to the subjective perception and experience of the operator, making it difficult to temporarily trace the state of the infrastructure and study itsrate of degradation.

Sewdefis a tool for digital transformation in the water sector

The SEWDEF system is a web application processed in a cloud environment, accessible from anywhere and through any device with an Internet connection.It bases its operation on the combination of computer vision algorithms, Artificial Intelligence (Deep Learning) and mobile robotics.

In this way, infrastructure managers will be able to obtain objective information on the state of the sewerage network and its temporal evolution in order to know its rate of degradation in order to plan and prioritize preventive maintenance actions.