If AI-based systems are to interact safely with humans and their environment they must be able to reliably recognize their surroundings.

Computer Vision

Intelligent systems must be able to reliably recognize people and objects from visual data taken from images or videos as well as develop an understanding of complex environments. This is a basic requirement if vehicles and mobile robots are to interact with their surroundings. Applying Artificial Intelligence to condition monitoring and quality control systems poses high demands on visual information processing.

We have many years experience of Machine Learning and image processing which gives us a competitive edge. Our systems have already been successfully put to the test in a range of applications and sectors and have proven themselves to work: For example, we offer solutions for recognizing road traffic signs, for condition monitoring systems in infrastructures, for damage detection and for quality analysis.

Our own core technology for image processing and neural network frameworks means we can recognize objects and analyze the surrounding area in real time. Data originating from different sources often needs to be combined and processed on distributed devices if a complex environment such as a specific traffic situation is to be analyzed.

These collaborative systems pose specific demands on resource efficiency and require special techniques for the fusion of data from different sensors. That is why we are working on and perfecting stand-alone Edge Computing systems and techniques which enable objects to be tracked through several fields of vision without involving a central cloud.

Traditional Machine Learning processes are often unable to recognize errors and quality defects in industrial production. By their very nature, errors and problems which only occur rarely can only be compared to a relatively small database. For application scenarios such as this we can develop hybrid processes and models which incorporate domain-specific expertise and can therefore also be trained to be used with smaller data sets.

Portfolio of services

Our customers cover the automotive, traffic, logistics, infrastructure and media sectors. Our image recognition and analysis systems help us analyze not only people and objects but also errors and quality defects.

Data assessment and improvement

We assess the reliability of image recognition, compensate for distortions and use data augmentation techniques to improve the base data.

Object recognition

We develop applications used for real-time recognition in videos of a variety of objects (such as road traffic signs) and situations (damages). Our facial recognition technology makes it possible to search video archives for specific individuals provided it is in the public interest and legal to do so.

3D semantic segmentation

3D semantic segmentation enables users to recognize and associate structures connected with each other in a particular environment.

Environment detection

Using the same techniques that enable environment detection in road traffic we are developing applications which also allow unusual environments such as construction sites, warehouses and buildings to be navigated safely.

Decentralized and collaborative image analysis

The neural networks we develop for applications on mobile devices and in collaborative systems are set to run using the least possible amount of resources. We also merge data derived from multiple sensors with the aim of using it in stand-alone Edge Computing systems.

Highlights

Image based damage detection in sewer systems

Many sewer networks are in need of repair. Because of the high level of manual labour required to fix them inspecting them can be a slow process. The automated detection of sewer pipe defects only makes economic sense if defects and damage can be detected with a sufficiently high degree of accuracy. As part of a project set up for the Berliner Wasserbetriebe water utilities company we developed algorithms to not only detect sewer damage but also to classify it according to type and severity.

Recognizing road traffic signs in roadworks

Our technology recognizes different highway code symbols regardless of whether the road traffic signs are round, triangular, rectangular or octagonal. Where it gets challenging, however, is when a road traffic sign contains specific information: such as when different speed limits are enforced at certain times of the day to control traffic noise or when complex lane guidance signs are in operation. By combining image and text recognition our solution reads the texts and understands the way signs are grouped together. This means drivers and navigation systems receive precisely the information that is relevant at that very moment.

Condition monitoring in rail infrastructures

Fallen trees on the tracks are a particularly high risk for the more than 33,000 kilometer long Deutsche Bahn rail network. Working together with reputable partners from the rail industry and research sectors we are developing 3D technology to reconstruct and analyze the rail network. Our software automatically recognizes relevant objects and works out to what extent they deviate from their target values. The analytical results will be visualized using one of our partner companies’ applications allowing infrastructure operators to plan and set in motion the necessary maintenance processes.

Image-based structural damage recognition

Machine-based learning processes being used to automatically recognize structural damage or errors quickly reach their limit if applied as part of industrial quality assurance with only a limited amount of suitable training data available. By incorporating domain-specific expertise we as part of the Fraunhofer Research Center for Machine Learning have successfully developed processes to train powerful Deep Learning systems to also use smaller data sets – such as for the automatic recognition of hail stone damage on vehicles or of material defects on smooth surfaces within industrial production.

Virtual perimeter advertising

Major events often use LED perimeter advertising boards showing electronically controlled changing animated visual content. Every TV audience around the world sees the same advertising because it has not been possible up until now to show country-specific advertising when broadcasting live, at least not in a satisfactory quality. With the system we have developed it is now possible to change the LED perimeter advertising content in real time: Any number of TV streams can now be generated showing different virtual advertising. This means that the TV audience does not realize that the perimeter advertising content has been changed on their TV screen. This system makes it possible to create completely new business models for marketing advertising rights at sports events.