基于图像的道路语义分割检测方法

Road semantic segmentation detection method based on image

针对现有语义分割网络模型在道路语义分割方面检测精度低、计算量大等问题,基于BiSeNet V2网络模型进行优化改进,引入一种高效的通道注意力(efficient channel attention,ECA)模块,在BiSeNet V2的语义分支和细节分支的每个阶段末端分别加入ECA,得到ECA-Semantic-BiSeNet V2网络。使用实车采集道路图像数据进行标注并构建自采数据集,在Cityscapes数据集、KITTI数据集及自采数据集上分别对改进前后的网络模型进行试验验证。试验结果表明,与BiSeNet V2模型方法相比,本研究方法在Cityscapes数据集上MIoU提高14.01%,在KITTI数据集上MIoU提高1.86%,同时在BiSeNet V2的语义分支加入ECA后运算量增加0.02 GFlops的条件下,模型推理速度提高了7.82帧/s。

To address the problems of low detection accuracy and high computational effort of the existing semantic segmentation net-work model in road semantic segmentation,this study optimized and improved the BiSeNet V2 network model by introducing an effi-cient channel attention(ECA)module.The ECA module was added at the end of each semantic path and detail path of BiSeNet V2 network respectively to obtain the ECA-Semantic-BiSeNet V2 network.Road image data collected by real vehicles were used to label and build self-collected data sets.The network models before and after improvement were tested and verified on Cityscapes data sets,KITTI data sets and self-collected data sets.The experimental results showed that compared with the BiSeNet V2 model method,this research method improved Miou by 14.01%on the Cityscapes dataset and 1.86%on the KITTI dataset,and the model reasoning speed increased by 7.82 fonts/s with the addition of ECA to the semantic path of BiSeNet V2 network with an increase of 0.02 GFlops.