Research On Key Technologies Of Semantic Segmentation In Road Scene Based On Deep Learning

In recent years,with the improvement of data computing capabilities,artificial intelligence has become more closely related to people’s lives.Advanced driver assistance systems have become the focus of machine intelligence and the automotive industry.Computer vision technology-based driver assistance systems can implement road environment perception,route planning,and driving decisions.As an important part of environmental perception,image semantic segmentation technology in road scenes can effectively improve the accuracy and speed of vehicle perception systems and enhance the capabilities of autonomous driving systems,so it has important research significance.Machine learning technology based on deep learning has made great progress in many fields,and deep learning has attracted much attention in the field of computer vision.The research purpose of this article is to study the data in the road scene based on deep learning,and design and implement an image semantic segmentation network that can meet the accuracy and real-time requirements of the assisted driving system.The main contents and innovations of the work are as follows:Based on thinking about the accuracy of semantic segmentation,we redesigned the segmentation network with better real-time performance,enhanced the capability of fine-grained feature extraction and multi-scale feature fusion,and improved the segmentation accuracy of network by using the complementation of low-level location information and high-level context information.At the same time,while ensuring accuracy,the network calculation is reduced,and a lightweight convolution operation is used to meet the real-time requirements.In the end,a segmentation network with high accuracy of semantic segmentation and inference speed meeting general requirements is realized.Based on the real-time requirements of semantic segmentation,the network structure with excellent real-time performance was modified.Utilizing the current design ideas of lightweight networks in other fields,we have designed lightweight feature extraction network,lightweight long-distance semantic fusion module and lightweight multi-scale feature fusion module which based on depth wise separable convolutions.The new module improves the speed of network inference without any loss of accuracy.Finally,a fast semantic segmentation network with good segmentation accuracy and high inference speed can be used.The semantic segmentation algorithm proposed in this paper can improve the segmentation accuracy and inference speed on the public data set,and has a good balance between the two,which satisfies the needs of driver assistance systems and contributes to research work in related fields.Some new methods and ideas have been introduced.