Front-to-Bird’s-Eye-View Transformation for Autonomous Vehicles: A Class Imbalance-Based Approach

Developing efficient and safe autonomous navigation systems for self-driving vehicles requires an accurate representation of the environment. Bird’s Eye View (BEV) provides a top-down view environment representation valuable for decision-making and path-planning. In this study, we explore the use of deep learning-based models to output BEV representations directly from front-perspective images, addressing challenges such as class imbalance. We design a synthetic dataset, test various loss functions, and propose a layer-based data augmentation strategy. The experimentation results and the discussion cover implemented deep learning models, data augmentation, and class weighting methods for loss training functions.