摘要
针对室内机器人行驶场景中存在大量不规则形状障碍物及细小类障碍物的问题,提出一种基于改进双边分割网络(BiSeNet)的图像语义分割方法.即以BiSeNet为基础网络构建图像分割模型,一方面在其空间路径中融合可变形卷积,使其更加适应对可通行区域和水渍这类不规则形状目标的定位分割;另一方面在其语义路径中结合特征金字塔结构,提高对细小类障碍物的分割精度;最后在室内多类障碍物图像数据集上,将改进的BiSeNet算法与U-Net,PSPNet等算法进行对比实验,结果显示改进的BiSeNet算法对水渍类障碍物的分割像素准确率达到89.95%,比原BiSeNet算法分割精度提高约3.50%,与UNet,PSPNet等算法相比,改进的BiSeNet算法同样具有更高的分割精度.
To solve the problems that there are a large number of irregular obstacles and small obstacles in the indoor robot driving scene,an semantic segmentation technology of images based on improved BiSeNet(bilateral segmentation network) was proposed.On the one hand,deformable convolution was integrated in its spatial path to make the model more suitable for flexible targets such as passable areas and water stains.On the other hand,feature pyramid structure was combined in its semantic path to improve the segmentation accuracy of small obstacles. Finally,the improved BiSeNet algorithm was compared with U-Net,PSPNet and other algorithms,on the image dataset of indoor multi-type obstacle. The results show that the improved BiSeNet algorithm has a pixel segmentation accuracy of 89.95% for water stains obstacles. Compared with the original BiSeNet algorithm,the segmentation accuracy is improved by about 3.50%. Compared with U-Net,PSPNet and other algorithms,the improved BiSeNet algorithm also has higher segmentation accuracy.
作者
于蒙
樊成
李雄
李文锋
YU Meng;FAN Cheng;LI Xiong;LI Wenfeng(School of Transportation and Logistics Engineering,Wuhan University of Technology,Wuhan 430063,China;Wuhan Mindray Medical Technology Research Institute Co.Ltd.,Wuhan 430075,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2022年第6期133-138,共6页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家重点研发计划资助项目(2020YFB1712400)
国家自然科学基金资助项目(71672137)。
关键词
室内障碍物图像
双边分割网络
语义分割
可变形卷积
特征金字塔网络
indoor obstacles image
bilateral segmentation network
semantic segmentation
deformable convolution
feature pyramid network
