关键词: 语义分割, 道路场景, 雨滴遮蔽, 类别权重, 移动端部署

Abstract: The existing semantic segmentation models are susceptible to the interference of raindrop occlusion， and the performance is poor on the rainy road scene dataset. Moreover， they did not focus on the two important categories of vehicles and pedestrians in the road scene. Aiming at the above two problems， this paper designs a semantic segmentation algorithm for rainy road scenes and deploys it on a mobile terminal to promote the development of autonomous driving technology. A fast fusion pyramid pooling module is proposed to make the feature map integrate rich global semantic information and local detail information， and effectively segment the raindrop obscured scene. A multiple attention fusion module is proposed， and the category weight of vehicles and pedestrians in the loss function is increased to enhance the model’s attention to vehicles and pedestrians. The model is deployed to the mobile terminal with the help of Android Studio platform， and the ONNX Runtime is used for forward inference， and the segmentation effect is consistent with that of the computer terminal. Compared with five recent models on the Rainy WCity dataset， the segmentation accuracy of this model is the same on the computer terminal and the mobile terminal. Specifically， PA and mIoU are 95.25% and 72.96%， vehicle PA and IoU are 84.04% and 74.15%， and pedestrian PA and IoU are 34.91% and 26.37%， respectively， which are higher than those of the other five models. In addition， the FPS of the model in the computer and mobile terminals are 45.46 and 1.26， respectively， and the segmentation speed is fast. The model proposed in this paper can effectively segment the road scene image under the shelter of rain on the mobile terminal， and it is more accurate to segment vehicles and pedestrians.

Key words:  , semantic segmentation； road scene； raindrop masking； class weight； mobile deployment