CD Laboratory for Semantic 3D Computer Vision

Robots are penetrating more and more areas of our lives. The aim of this CD Laboratory is to enable robots to orientate themselves independently in a non-standardised environment and perform certain tasks.

Computer-controlled image recognition enables a robot to recognise where it is, which object it should handle and where it is located relative to itself. Not least due to insufficient computing power, attempts have so far been made to describe the environment to robots on the basis of two-dimensional images and there is a great deal of work on standardising the description of visual information. The problem, however, is that the underlying images can be subject to numerous disturbances or deviations that make it difficult for the robot to process them independently. In the case of outdoor images, the weather, time of day or season can vary, images can be sharp or blurred, target objects can be very similar and therefore difficult to describe specifically and sometimes the target is the empty space between numerous objects, which is difficult to capture descriptively. This is why robots have so far mainly proved their worth in a controlled environment. However, sufficient computing power is now available to enable image description in 3D and thus autonomous environment recognition by robots in a real, uncontrolled environment.

One of the most promising applications of computer vision is robot-assisted maintenance in factories. Position recognition in outdoor areas could also far exceed the accuracy of GPS. Currently, such applications are laboriously programmed by hand or learnt by the robot through repetition and are very error-prone and not very flexible. Statistical methods, which are already being used in big data and machine learning, could offer a powerful option here. However, machine learning techniques are not yet capable of solving tasks such as determining the position of an object in 3D. The aim of this CD Laboratory is to close the gap between machine learning techniques and geometric computer vision. Statistical algorithms are to be developed as fundamental building blocks for 3D computer vision applications. The aim is to ensure a sufficient balance between the accuracy of the results and the speed of their calculation.

The research work will focus on various camera types and sensors for environment and position recognition, as well as motion and acceleration sensors and compasses, in order to provide the robot with a diverse data source for determining its position, even at high speeds. Furthermore, ways are to be found to enable automated position recognition based on available image databases, such as Google Street View, or by using two-dimensional sources, such as maps and city plans.

The aim is to ensure that robots and machine learning techniques can leave the laboratory and realise their many possibilities in an uncontrolled environment.