Minimizing Supervision for Vision-Based Perception and Control in Autonomous Driving

The research presented in this dissertation focuses on reducing the need for supervision in two tasks related to autonomous driving: end-to-end steering and free space segmentation.

For end-to-end steering, we devise a new regularization technique which relies on pixel-relevance heatmaps to force the steering model to focus on lane markings. This improves performance across a variety of offline metrics. In relation to this work, we publicly release the RoboBus dataset, which consists of extensive driving data recorded using a commercial bus on a cross-border public transport route on the Luxembourgish-French border.

We also tackle pseudo-supervised free space segmentation from three different angles: (1) we propose a Stochastic Co-Teaching training scheme that explicitly attempts to filter out the noise in pseudo-labels, (2) we study the impact of self-training and of different data augmentation techniques, (3) we devise a novel pseudo-label generation method based on road plane distance estimation from approximate depth maps.

Finally, we investigate semi-supervised free space estimation and find that combining our techniques with a restricted subset of labeled samples results in substantial improvements in IoU, Precision and Recall.