Research On Semantic Segmentation Technology Of Unmanned Vehicle Image Based On Multi-scale Feature Fusion

With the rapid development of computer science and technology,the relationship between artificial intelligence and logistics industry is becoming closer and closer.The traditional time-consuming and laborious distribution mode can not meet the needs of people’s fast life,so it puts forward new requirements for the field of logistics distribution.In order to further optimize the end distribution service,improve the efficiency of logistics distribution and reduce the logistics cost,unmanned distribution has become the focus of researchers.The realization of unmanned distribution depends on professional driverless system.How to correctly understand the road scene information is the key of driverless system and the basis of autonomous path planning of unmanned distribution vehicles.The application of image segmentation technology in driverless system can effectively improve the accuracy and speed of road condition judgment,so it has very important research significance.In recent years,the performance of deep learning in the field of image segmentation has made great progress,which has exceeded the traditional image segmentation algorithms.Therefore,this paper studies image segmentation in road environment based on deep learning,and designs a semantic segmentation network to meet the requirements of driverless system with the goal of accuracy and real-time,so as to provide technical support for accurate scene understanding and analysis of subsequent systems.The specific research results are as follows:(1)A multi-scale semantic segmentation network for road scene is proposed.Based on the accuracy requirements of the semantic segmentation model for the road scene,a semantic segmentation network based on multi-scale feature fusion is constructed.The network can help the driving system of unmanned distribution vehicle identify the road conditions more accurately and provide technical support for the subsequent safe driving.The network model includes two paths.The spatial path uses deformable convolution to learn shallow spatial detail features,and the semantic path uses void convolution to learn multi-scale advanced semantic features.The segmentation accuracy of the network is improved through the complementarity of spatial information and semantic information.The segmentation accuracy of the proposed method on Cityscapes data set is improved by 1.2%compared with the comparison network.(2)A multi-scale lightweight semantic segmentation network for road scene is proposed.Considering the real-time requirements of the driving system,a lightweight semantic segmentation network with both accuracy and real-time is constructed,which can help the driverless system realize the demand of real-time distribution.The network model includes spatial path and semantic path.The deep separable convolution is used to reduce the amount of network parameters and computation,speed up the segmentation speed,and fuse the characteristics of the two paths is the final segmentation result.The proposed method achieves the segmentation accuracy of 68.4%Miou at the speed of 85.9 frames per second on the Cityscapes data set,which has certain advantages compared with the existing mainstream models.(3)This paper designs a semantic segmentation algorithm verification system for road scene,and processes the above algorithm model in engineering.The system is built based on python programming language and embedded with the above lightweight semantic segmentation network.Simulating ordinary users to test the verification system,it proves the effectiveness of semantic segmentation network,which makes the algorithm studied in this paper have certain application significance,and the system has the characteristics of simple operation,high modularity and low cost.