Semantic Segmentation Of Road Scene Based On Deep Learning

With the development of deep learning,autonomous driving has gradually become a research hot spot.An important part of autonomous driving is the perception of the driving environment of the car,and the semantic segmentation of road scene based on deep learning is one of the most commonly used methods in the environment perception.Therefore,this paper focuses on the semantic segmentation of road scenes based on deep learning,and elaborates on the theory of deep learning and image semantic segmentation.This paper proposes two semantic segmentation networks for road scene based on deep learning.The first one is attention-guided residual network for semantic segmentation that can achieve competitive results without pretrained models and backend processing.The second one is a dual-stream multi scale network for real-time semantic segmentation,which is excellent in both accuracy and speed.The specific work as follows:(1)Attention-guided residual networks for semantic segmentation: The fullresolution residual unit proposed by FRRN has no feature selection capability.In response to this problem,this paper introduces the attention mechanism and proposes the attention-guided residual unit.In addition,we analyzed that attention mechanism will continue decrease the value of the feature map,and the attention mechanism cannot be stacked in the network.In order to solve this problem,we designed a stackable attention-guided residual unit from residual attention module.At last,we proposed a attention-guided residual networks for semantic segmentation.Without the pretrained model and any backend processing,our network achieved 62.9% MIoU of the Cityscapes dataset,got an increase of 2 percentage points compared to FRRN.(2)Dual-stream multi scale network for real-time semantic segmentation: Road scene semantic segmentation has real-time requirements.For this reason,we analyzed the factors that affect the accuracy and speed of semantic segmentation.We found that in the most semantic segmentation networks structure,spatial information is gradually lost as the network acquires advanced context features.And using the deeper networks to obtain advanced context information will reduces the speed of the model.Based on the above observation,we propose a network for semantic segmentation with spatial information stream and context information stream.In order to further improve the accuracy,we combined the multi scale features of two streams through a feature fusion module with attention mechanism.Therefore,we refers to it as a dual-stream multi scale network for real-time semantic segmentation.On the CamVid dataset,our network used ResNet18 as backbone network MPA reached 74.8%,MIoU reached 58.4%,and the speed reached amazing 289 fps.