Abstract

Foreground moving object segmentation is a critical challenge in various computer vision applications. Background modeling techniques have made significant progress, but achieving accurate foreground segmentation remains elusive. Most existing algorithms operate exclusively within the color space, making them susceptible to issues like lighting changes, shadows, automatic camera adjustments, and color camouflage. Obtaining large-scale datasets with hand-labeled ground truth is both time-consuming and labor-intensive, rendering these methods challenging to implement in practice. This work presents a self-supervised learning solution for drivable area and road anomaly segmentation, bypassing the need for manual labeling. The proposed method automates the generation of segmentation labels for drivable areas and road anomalies. It leverages RGB-D data to train neural networks for semantic segmentation, using a pipeline known as the Self-Supervised Label Generator (SSLG) to create segmentation labels. The SSLG-generated labels are then employed to train multiple RGB-D data-based semantic segmentation neural networks.

Description