基于Transformer的道路场景语义分割综述  

Review of semantic segmentation of road scene based on Transformer

作　　者：黄天云[1] 向明建 邵世霖 HUANG Tianyun;XIANG Mingjian;SHAO Shilin(School of Mathematics,Southwest Minzu University,Chengdu 610225,China;School of Computer and Artificial Intelligence,Southwest Minzu University,Chengdu 610225,China)

机构地区：[1]西南民族大学数学学院,四川成都610225 [2]西南民族大学计算机与人工智能学院,四川成都610225

出　　处：《西南民族大学学报(自然科学版)》2025年第2期193-205,共13页Journal of Southwest Minzu University(Natural Science Edition)

基　　金：中央高校基本科研业务费专项资金研究生创新项目(YCYB2024001)。

摘　　要：在自动驾驶领域,通过对道路场景进行高质量的语义分割,可以为自动驾驶汽车的安全行驶提供重要保障.近年来,随着自动驾驶技术的不断进步,人们对语义分割模型在尺寸、计算成本和分割精度等方面的要求也日益提高,这促使研究者们探索更为先进的算法.首先介绍了语义分割技术在深度学习快速发展下取得的显著进展与不足,从而引出基于Transformer的道路场景语义分割方法.相较于传统的深度学习算法,Transformer具备全面理解复杂场景中上下文关系的能力,尤其在处理多对象和复杂环境时表现出显著优势.接着,根据不同的特征处理策略和模型架构,将基于Transformer的道路场景语义分割方法分为四类:基于全局特征提取的方法、基于局部特征增强的方法、基于混合架构的方法以及基于自监督学习的方法.最后,分析和对比了每类方法的代表性算法,概括总结了各类方法的技术特点和优缺点.In the field of autonomous driving,high-quality semantic segmentation of road scenes is crucial for ensuring the safe operation of self-driving cars.In recent years,with the continuous improvement of autonomous driving technology,the demand has been increasing for semantic segmentation models in terms of size,computational cost,and segmentation accuracy,prompting researchers to explore more advanced algorithms.This paper first introduced the significant progress and shortcomings of semantic segmentation technologies under the rapid development of deep learning,leading to the discussion of Transformer-based methods for road scene semantic segmentation.Compared to traditional deep learning algorithms,Transformers possessed the ability to comprehensively understand contextual relationships in complex scenes,demonstrating significant advantages,particularly in handling multi-objects and complex environments.Subsequently,based on different feature processing strategies and model architectures,the Transformer-based road scene semantic segmentation methods were categorized into four types:methods based on global feature extraction,methods based on local feature enhancement,methods based on hybrid architectures,and methods based on self-supervised learning.Finally,the paper analyzed and compared representative algorithms from each category,summarizing their technical characteristics,advantages and disadvantages.

关 键 词：语义分割 TRANSFORMER 全局特征提取 局部特征增强 混合架构 自监督学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]