Exploring diffusion models for semantic segmentation in bird's eye view mapping for autonomous vehicle perception

Abstract

This thesis, investigates the application of diffusion models to improve the perception systems of autonomous vehicles. Leveraging semantic segmentation, we utilize bird's eye view mapping as a novel perspective for more robust environment comprehension. With a specific focus on the widely recognized NuScenes dataset, the research methodically unpacks the process of image segmentation, the propagation of error through the model, and the subsequent impacts on system comprehension and prediction. Emphasis is placed on the potential enhancements offered by diffusion models, as they allow for better uncertainty quantification, a key aspect in safety-critical applications such as autonomous driving. While the results achieved did not surpass the current state-of-the-art (SOTA) models, they provide critical insights into the nature and potential improvements in the handling of segmentation tasks for autonomous vehicles. The insights and learnings from this study will contribute significantly to future research aiming to refine and optimize perception models for autonomous driving.