George Basem Fouad Eskandar

Autonomous Driving systems are redefining the role of cars and reshaping the future of urban transportation. In recent years, the breakthrough in deep learning technology and computer vision has catalyzed the research in Autonomous Driving. Although significant progress has been achieved, numerous obstacles remain the path of robust Autonomous Driving. Deep Learning models require large datasets for training, which do not always involve all environmental settings that will be encountered at test time. Unseen weather and lighting conditions as well as unseen street layouts in different countries can affect the performance of perception systems in the self-driving car.

To this end, a prominent line of research has emerged in Computer Vision, namely Domain Adaptation (DA), that seeks to address this problem. DA leverages labeled data in one or more “source domains” to train a model that achieves a good accuracy on a “target domain”. Usually source and target domain distributions are different from each other.

My research focuses on exploring many Domain Adaptation techniques in Autonomous Driving. More specifically, my research includes the following topics:

• Developing self-supervised and unsupervised Generative Adversarial Networks, to be used as a data augmentation tools for Domain Adaptation.

• Developing new applications of GANs on LiDAR data to allow for a more robust driving. LiDAR is a 3D laser scanning sensor technique that is crucial in Autonomous Driving.

• Exploiting several DA techniques in the perception systems of self-driving cars with a special focus on sensor fusion

More details about the research project I work with can be found here: https://www.ki-deltalearning.de/