Semantic Segmentation of Urban Street Scenes Based on Improved DeepLabv3+

Abstract

Intelligent transportation and autonomous driving systems require precise semantic segmentation of urban street scenes. The goal of this work is to enhance pixel-level semantic segmentation performance in challenging urban settings that typically have issues like class imbalance, multi-scale context, and fine object boundaries. A high-end segmentation system is shown that enhances the DeepLabv3+ backbone with an adaptive multi-scale context aggregation module, an edge-aware refinement module, and a context attention method. The Cityscapes and CamVid urban scene datasets have been used in numerous projects. The suggested approach outperformed strong baselines by 2.4% and 1.2% on the Cityscapes test set, with mean Intersection-over-Union (mIoU) and pixel accuracy of 83.7% and 97.1%, respectively. Additionally, there has been a notable improvement in segmentation accuracy for the small and thin class of poles and riders. Qualitative visualization also demonstrates improved boundary delineation and occlusion robustness in a variety of real-world circumstances. According to the aforementioned findings, the new architecture can enhance the precision and consistency of semantic segmentation for challenging urban scenarios, offering a more reliable foundation for the development of intelligent visual systems.