Research And Implementation Of Image Semantic Segmentation Based On Autonomous Driving

In recent years,with the frequent occurrence of traffic safety accidents and traffic congestion problems such as drunk driving and sudden accidents,more attention has been paid to autonomous vehicles and their applications in the market.Among them,sensing a large amount of external environment information can ensure the correct planning and decisionmaking of autonomous driving vehicles,and image semantic segmentation,as one of its key technologies,has become a research hotspot.In most scenes,image semantic segmentation algorithms pursue higher segmentation accuracy,so they use larger and deeper neural network structure,which will lead to the decline of network reasoning speed,longer algorithm reasoning time,and higher storage cost,which can’t meet the real-time requirements of autonomous driving applications.However,if only a simple network is adopted,the speed can be ensured to some extent,but the image segmentation effect may not be ideal enough to meet the requirements of segmentation accuracy.Therefore,the optimization of segmentation accuracy and inference speed is a major challenge for image semantic segmentation algorithms for autonomous driving.This paper discusses and studies the image semantic segmentation algorithms inspired by convolutional neural network(CNN),and further proposes an improved and optimized scheme to solve the problems of spatial detail feature loss and insufficient feature extraction.Therefore,the main contents of this paper are as below:(1)According to the analysis of the existing network structure,codec structure is a good way to extract features,and two-branch structure can speed up the network.In order to achieve the relative balance between the segmentation accuracy and reasoning speed,this paper takes Fast-SCNN,which combines the advantages of the two structures,as the basic network structure for subsequent improvement.(2)In order to speed up the processing speed of the network,a new convolution method and lightweight residual module are adopted in this paper,which greatly reduce the number of parameters of the model,making the model occupy a low amout of memory and easy to be embedded in the limited mobile devices such as self-driving system.(3)In order to improve the segmentation accuracy of the model,this paper designs a finegrained hierarchical residual structure to extract features,and then analyzes the impact of different global feature extractors on the network performance through comparative experiments,thus verifying the advantages of this structure.In addition,a new learning down-sampling improvement scheme is proposed,which combines with VGG network convolution layer stacking mode to enhance the ability of spatial detail feature extraction.Finally,the effectiveness of the algorithm is verified by experimental analysis.(4)In order to further improve the algorithm performance,the bar pooling module SPM is integrated into the pyramid pooling module PPM to reduce the disadvantage that the traditional spatial pooling method cannot be associated with long-range pixels,and at the same time,the large spatial pooling core is removed to improve the efficiency of the algorithm.The optimization scheme is compared from multiple dimensions to select the algorithm with the best performance.The improved algorithm is analyzed by qualitative index and uncertainty partition difference graph,and the comparison experiment is made with the state-of-the-art algorithms on Cityscapes dataset.The results show that the improved algorithm reaches 73.4% in the average segmentation accuracy MIo U,which is 5.2% higher than the original algorithm,and achieves the processing speed of 82.7 frames per second,which to some extent realizes the requirements of the autonomous driving system for high-precision,real-time and lightweight image semantic segmentation algorithm.