PhD defence by Kashita Niranjan Udayanga

Abstract

Civil infrastructure such as roads, bridges, and buildings—requires regular inspections to ensure public safety. If defects like cracks or structural damage go unnoticed, they can lead to serious failures or even collapses, putting lives at risk. Traditionally, these inspections are done by human workers, which can be slow, costly, and sometimes dangerous. Especially when inspecting hard-to-reach or hazardous areas. Aerial robots, or drones, offer a promising alternative. Equipped with cameras, they can quickly and safely capture images of infrastructure, even in risky environments. However, these inspections generate thousands of images, making manual analysis slow, error-prone, and inefficient. That’s where artificial intelligence - specifically, deep learning— comes in.

By training AI models to detect defects automatically, we can significantly speed up and improve the accuracy of inspections. However, this approach still faces real-world challenges. First, lighting conditions: drones often operate in difficult lighting - at night, in shadows, or under harsh sunlight. Standard cameras struggle in these scenarios. In our project, we explored using an advanced type of sensor called a Dynamic Vision Sensor (DVS), which can handle extreme lighting conditions with minimal need for extra lighting. Second, energy efficiency: whether processing happens onboard the drone or at the inspection site, power is limited. specially during real-time, on-site inspections. Running deep learning models on traditional processors can quickly drain batteries and require bulky hardware.

To address this, we investigated neuromorphic processors, an emerging type of chip that mimics how the human brain works - using Spiking Neural Networks (SNNs). These processors, like Intel’s Loihi2, are designed to be ultra energy-efficient, making them ideal for edge devices like drones. Our project brings together three cutting-edge technologies: Dynamic Vision Sensors (DVS), Spiking Neural Networks (SNNs) and Neuromorphic Processors. All these falls under the exciting field of neuromorphic computing. We explored how this combination can enable real-time, energy-efficient defect detection during civil infrastructure inspections.

Supervisors

• Main supervisor: Associate Professor Silvia Tolu, Department of Electrical and Photonics Engineering, DTU
• Co-supervisor: Associate Professor Matteo Fumagalli, Department of Electrical and Photonics Engineering, DTU
• Co-supervisor: Dr. Cesar Dario Cadena Lerma, Department of Mechanical and Process Engineering, ETH Zürich

External examiners

• Associate Professor Francisco Barranco, Computer Engineering, Automation and Robotics, University of Granada, Spain
• Professor Andrea Acquaviva, Electrical, Electronics and information engineering, University of Bolonia, Italy

Chair of assessment committee

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

Master of the Ceremony

• Associate Professor XinXin Zhang, Department of Electrical and Photonics Engineering, DTU