Scenario-Based Segmentation: Traffic Image Segmentation by GNN Based Driver's Scenario

Abstract

This paper introduces the Scenario-Based Segmentation Network (SBS-Net), which highlights significant advances in autonomous driving. Through the integration of the Scenario Enhanced Graph Neural Network (SE-GNN) and graph re-match modules into the existing semantic segmentation network model based on driver's cognition (DCSeg-Net), our approach optimizes the graph construction process. The SE-GNN module enhances the extraction of critical features and relations within diverse driving scenarios, elevating both accuracy and the system's adaptability to complex driving scenarios. The introduced graph re-match module refines classification discrepancies, significantly boosting segmentation accuracy and refining the understanding of autonomous driving scenes. Beyond technological enhancements, this work outlines the expansion and diversification of our original dataset, strengthening the learning capabilities of our model by including specific classification labels and incorporating a broader range of driving scenarios. The utility of the improved SBS-Net is demonstrated through superior performance in Graph Construction Accuracy and Intersection over Union (IoU) measures, as highlighted in our evaluation metrics. These advances underscore the practical applicability of scenario-based segmentation in real-world autonomous driving scenarios, enhancing overall scene comprehension capabilities. The developments presented signal substantial progress in the field of autonomous driving technology.