Research On Semantic Segmentation Algorithm Of Road Scene Based On Deep Learning

With the rapid development of related technologies such as computer vision,autonomous vehicles have been launched and updated quickly.Accurate perception of the surrounding environment is an extremely basic and important part of autonomous driving technology.Image semantic segmentation technology can segment the road scene image acquired by the car into regions representing different categories,and can provide an important basis for autonomous vehicles to perceive the surrounding environment and make decisions.It has become an indispensable technology in autonomous driving tasks.As road scene images have the characteristics of a wide variety of targets and large scale changes,semantic segmentation of road scenes has become a challenging problem.Traditional image segmentation methods rely heavily on manual feature analysis and feature extraction,which is not only arduous,but also difficult to find comprehensive features,which makes it difficult to effectively segment images in complex scenes.In recent years,deep learning has developed rapidly.It can automatically learn rich deep features from image data,and obtain high-precision semantic segmentation results based on the extracted features.The road scene semantic segmentation method based on deep learning overcomes the limitations of traditional methods and has high research value.At present,although the deep learning methods applied to the semantic segmentation of road scenes have achieved certain results,there are still problems such as low segmentation accuracy and insufficient segmentation speed.Based on the above problems,this paper deeply researches the road scene semantic segmentation algorithm based on deep learning.The specific research content is as follows:1.A road scene semantic segmentation algorithm that simulates the brain's visual mechanism was proposed.(1)Two branches were designed to simulate the ventral and dorsal pathways of the visual cortex of the brain.The ventral pathway network focuses on extracting the semantic information of the image,and the dorsal pathway network focuses on extracting the spatial information of the image,and the U-shaped structure was combined to obtain richer spatial information.Finally,the information of the two pathways of were fully integrated to obtain high-precision segmentation results.(2)The semantic enhancement module and the spatial attention module were proposed to strengthen the extraction of semantic information from the ventral pathway network and the extraction of spatial information from the dorsal pathway network,respectively.The semantic enhancement module was improved by integrating the channel attention mechanism in atrous spatial pyramid pooling,which can focus on fusing semantic information with different scales.The spatial attention module enabled the network to focus on more important information such as the target edge in the spatial dimension.(3)By fine-tuning the structure of this algorithm model,it was lightened to meet real-time requirements while maintaining high segmentation accuracy.(4)Experiments showed that the algorithm achieved 82.1% segmentation accuracy;the lightweight model achieved 75.9% segmentation accuracy and106.9FPS inference speed,achieving a balance between accuracy and speed.2.A road scene semantic segmentation algorithm based on dual input feature pyramid network was proposed from the perspective of feature fusion.(1)The ventral and dorsal pathways structure proposed in the previous algorithm was modified,and the data was received and processed by two pathways,focusing on the extraction of semantic information and spatial information respectively.This ensured that the model can not only extract rich information,but also run with fewer parameters.(2)A novel dual input feature pyramid structure was designed to receive features with different information from the above two pathways.And after multiple effective feature fusion,the final segmentation result was output.At the same time,the intermediate feature map was reused,thereby reducing the amount of calculation.(3)Experiments showed that the algorithm achieved a segmentation accuracy of76.9% and an inference speed of 107.0FPS,realized real-time semantic segmentation with high accuracy,and can be applied to road scene image analysis tasks in real environments.