Attila Herbert

A thesis project for Maastricht University on explanations for semantic segmentation models.

I developed a new method called Proportionality-driven Occlusion Grid (POG), which outperforms current state-of-the-art explanations. I researched and wrote all the code for my thesis project, which involved learning about semantic segmentation and its challenges.

Several German companies launched this project to share information and create a framework for safer AI techniques in autonomous driving. I was part of this project as a consultant on model robustness, which included researching methods to simulate adverse environments and train models on that simulation and augmenting training data for robustness, implementing these methods, and finally, testing if the resulting models are actually safer.

During this project, I had to work with tight deadlines, give presentations to researchers and high-ranking company officers, and respond to any request regarding model safety in extreme environments.

This project aims to detect any foreign objects in mined ore (natural rock or sediment) transported on conveyors using cameras placed above the conveyor. I was responsible for any AI-related software in this project. I analyzed the provided data to ensure its quality, developed software that trains an object detector, deployed trained models to Jetson devices using TensorRT, and researched methods for streaming results from the edge device to the mine's monitoring software.

Designed, prototyped, implemented, and deployed an application that tracks mining trucks in an open-pit mine. The trucks are detected, classified, and tracked. I also created a web app that displays the trucks where the user can select and zoom onto specific trucks.

I created the back end using YOLOv8, TensorRT, and FastAPI and the front end with React and Express.js. I also did the model training, including overseeing the annotation process.

The application was deployed to run in Docker on a Jetson Edge unit.