RGB2BEV-Net: A PyTorch-based End-to-End Pipeline for RGB to BEV Segmentation Using an Extended Dataset for Autonomous Driving

In this work, we introduce RGB2BEV-Net, an end-to-end pipeline that extends traditional BEV segmentation models by utilizing raw RGB images with Bird’s Eye View (BEV) generation. While previous work primarily focused on pre-segmented images to generate corresponding BEV maps, our approach expands this by collecting RGB images alongside their affiliated segmentation masks and BEV representations. This enables direct input of RGB camera sensors into the pipeline, reflecting real-world autonomous driving scenarios where RGB cameras are commonly used as sensors, rather than relying on pre-segmented images. Our model processes four RGB images through a segmentation layer before converting them into a segmented BEV, implemented in the PyTorch framework after being adapted from an original implementation that utilized a different framework. This adaptation was necessary to improve compatibility, optimize performance and ensure better integration of the entire system within autonomous vehicle applications. We evaluate the performance of RGB2BEV-Net on the newly extended dataset and present a detailed result analysis, compared to traditional pre-segmented approaches. The system demonstrates flexibility and potential applicability for use in next-generation ADAS and autonomous navigation systems.