基于热力图像的道路场景稠密多级语义分割方法

Road Scene Dense Fusion Semantic Segmentation Method Based on RGB-T

语义分割能帮助自动驾驶实现精准感知。为解决彩色图像在模糊或昏暗场景下分割效果差,提出一种基于彩色图像和热力图像的多级融合语义分割网络。将语义分割任务分解为4个阶段,包括稠密多级融合特征提取、前景显著定位、目标分割任务和边界细化。主干网络提取热红外信息,将提取后的信息作为辅助信息与彩色图像信息进行融合,对融合后的特征进行前景和背景区域的区分,并用于目标分割和边界细化。多场景实验仿真表明,本方法能够有效提升模糊和昏暗场景下的道路场景分割效果。

Semantic segmentation helps autonomous driving system achieve accurate perception.In order to solve the problem of poor segmentation effect based on fuzzy or dim RGB image,we propose a new semantic segmentation method based on RGB(Red Green Blue)image and thermal image.The method divides the semantic segmentation task into four stages,i.e.,dense fusion feature extraction,foreground salient location,object segmentation and boundary refinement.The backbone network first extracts thermal information,and then uses the extracted information as auxiliary information to fuse with RGB information.The model distinguishes the foreground and background areas of the fused features,and makes the fused features more focus on the foreground target through the attention mechanism module,which is used for object segmentation and boundary refinement.After a lot of experiments,it is proved that this method can effectively improve the segmentation effect of road scene in fuzzy and dim scenes.