基于鸟瞰视图的递归特征金字塔语义分割方法
Recursive Pyramid Semantic Segmentation Method Based on Autopilot Aerial View
高宏伟 1赵博杨1
作者信息
- 1. 沈阳理工大学 自动化与电气工程学院,沈阳 110159
- 折叠
摘要
针对现有自动驾驶场景的鸟瞰视图(BEV)语义分割方法的特征提取效果差、视角转换方法老旧以及交通环境语义分割效果不佳的问题,提出一种基于鸟瞰视图的递归金字塔的语义分割方法,简称为IRFPN.传统金字塔特征提取网络包含自底向上、自顶向下以及横向链接三部分,特征提取效果有待提高.鸟瞰视角转换方面只使用传统的逆透视变换处理车载相机拍摄的图片,不能对相关语义进行很好的保留.为改善上述不足,将递归特征金字塔结构和极射线映射相结合.实验结果表明,IRFPN在nuScenes数据集中常见类别语义分割精度提升了0.4%,对驾驶区域的预测精度提升了2.1%,相比传统的金字塔特征提取网络效果好,对各常见交通环境语义特征有更好的保留,对复杂的交通环境具有较好的鲁棒性.
Abstract
A recursive pyramid semantic segmentation method based on bird's-eye views(IRFPN)is proposed to address the issues of poor feature extraction,outdated perspective conversion meth-ods,and poor semantic segmentation performance of existing bird's-eye views(BEV)semantic seg-mentation methods for autonomous driving scenarios.The traditional feature pyramid extraction net-work includes three parts:bottom-up,top-down,and horizontal linking,and the effectiveness of fea-ture extraction needs to be improved.In terms of bird's-eye perspective conversion,the relevant se-mantics cannot be preserved wellonly by using traditional inverse perspective transformation to process images captured by car mounted cameras.To overcome these shortcomings,a recursive fea-ture pyramid structure and polar ray mapping are combined.The experimental results show that IR-FPN has improved the accuracy of common categories in the nuScenes dataset by 0.4%and the prediction accuracy of driving areas by 2.1%.It can be concluded that recursive feature pyramids are better than traditional pyramid feature extraction networks,with better retention of semantic fea-tures in various common traffic environments,and better robustness to complex traffic environ-ments.
关键词
递归特征金字塔/特征提取/鸟瞰视图/语义分割/自动驾驶/极射线映射Key words
recursive feature pyramid/feature extraction/aerial view/semantic segmentation/auto-pilot/polar ray mapping引用本文复制引用
基金项目
辽宁省科学技术计划(机器人学国家重点实验室联合开放基金)项目(2021-KF-12-05)
出版年
2024
NETL
NSTL
维普
万方数据