基于改进BiSeNet的实时道路场景分割算法  

Research on real-time road scene segmentation algorithmbased on improved BiSeNet

作　　者：宋宇[1] 张丽影 梁超[1] SONG Yu;ZHANG Liying;LIANG Chao(School of Computer Science&Engineering,Changchun University of Technology,Changchun 130102,China)

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130102

出　　处：《长春工业大学学报》2024年第5期412-420,共9页Journal of Changchun University of Technology

基　　金：吉林省科技发展计划项目(20220201030GX)。

摘　　要：无人驾驶场景中的图像信息具有目标种类多、场景变化多样的特点,使得分割任务更加困难。针对空间信息和语义信息直接进行特征融合效果较差,面对复杂交通场景时分割结果较为粗糙,对双分支网络BiSeNet进行优化改进。采用轻量级骨干网络以及引入混合深度卷积,降低运算成本的同时提升分割精度;对复杂道路场景图像分割时存在结果边缘模糊、小目标分割效果不佳等问题,在上下文路径引入坐标注意力机制获取更多的上下文信息,提高分割的准确度;重新设计特征融合模块,使用注意力机制驱动方法构建特征融合模块,缩小了空间信息和语义信息之间层级的差距。在CamVid数据集上对改进后的算法进行训练并测试,算法平均交并比达到69.9%。改进网络虽然分割速度稍有降低,但其训练结果的平均交并比相比BiSeNet提升了1.6%。The image information in unmanned scenes is characterized by many kinds of targets and varied scenes,which makes the segmentation task more difficult.In view of the poor effect of direct feature fusion of spatial and semantic information,and the rough segmentation results when facing complex traffic scenes,BiSeNet based on two-branch network is optimized and improved.By adopting a lightweight backbone network and introducing mixed depth convolution,the segmentation accuracy is improved while reducing the computing cost;for the problems of blurred edges and poor segmentation of small targets in image segmentation of complex road scenes,the coordinate attention mechanism is introduced in the context path to obtain more contextual information and improve the segmentation accuracy;the feature fusion module is redesigned,and the feature fusion module is constructed using the attention mechanism-driven method to narrow the size of the feature fusion module and reduce the size of the feature fusion module.The feature fusion module is redesigned to use the attention mechanism-driven method to construct the feature fusion module,which reduces the gap between the layers of spatial information and semantic information.The improved algorithm is trained and tested on the CamVid dataset,and the average cross ratio of the algorithm reaches 69.9%.Although the segmentation speed of the improved network is slightly reduced,the average crossover ratio of its training results is improved by 1.6%compared to BiSeNet.

关 键 词：无人驾驶 实时语义分割 深度学习 注意力机制 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]