PhD defence by Patrick Schmidt

Abstract

Autonomous systems like robots and self-driving machines are becoming more common, but most of them are designed for clean, predictable environments like warehouses or roads. What happens when we try to use them in messier, more complex places like farms or construction sites? That’s the challenge the Ph.D. thesis is tackling. The goal: to help robots see and understand their surroundings in tough environments especially in reinforced concrete construction, where conditions are chaotic. Not only this, but the data needed to build an AI model for an autonomous system for this task is scarce, and labelling images, which is mostly required to train an AI, is expensive and time-consuming.

To address this, in this Ph.D. thesis we presented research that combines traditional, non-AI-based methods that process camera images with those that use AI, combining the best of both worlds. We need lots of data to train an AI model, so we collected a dataset for reinforced concrete construction, ConRebSeg. This data helped us to teach robots to “see” exposed metal bars in concrete structures, which is an indicator for defective concrete.

As mentioned previously, labelling the images of this dataset was tedious. Because we deal with robotic applications of AI models, additional data is available that can help us to make this labelling process easier. Instead of drawing the outlines of objects in the image, we can just indicate whether an object is present in the current image or not and still get the outlines of the objects. We showed that the additional depth data that robotic applications carry, i.e. how far away objects are in an image, can help to improve these methods.

Overall, our research showed how to cleverly use AI methods when they are applied to robots, and especially how to deal with the shortage of data when training AI models.

Supervisors

• Main supervisor: Professor Lazaros Nalpantidis, Department of Electrical and Photonics Engineering, DTU
• Co-supervisor: Associate professor Evangelos Boukas, Department of Electrical and Photonics Engineering, DTU

External examiners

• Professor Antonios Gasteratos, Democritus University of Thrace, Greece
• Associate Professor Chen Li, Aalborg University, Denmark

Chair of assessment committee

• Associate Professor Dimitrios Papageorgiou, Department of Electrical and Photonics Engineering, DTU

Master of the Ceremony

• Associate Professor Søren Hansen, Department of Electrical and Photonics Engineering, DTU