Research On Cityscape Scene Parsing Method For Automatic Driving

Since the concept of autonomous driving was born,it has been an important technology that countries are competing to develop.However,current self-driving cars are still unable to accurately and efficiently analyze and recognize the surrounding environment,especially in complex urban streets.Scene parsing requires providing a complete understanding of the scene,which predicts the label and location of each pixel in the entire image,and classifies and localizes the objects it belongs to.There are two key difficulties in the analysis of complex urban landscape scenes,namely,the maintenance of high resolution and the acquisition of high-level semantic information.Based on deep learning technology,this paper proposes an improved high-resolution network combined with hollow space pyramid pooling and a highresolution scene analysis network for two major difficulties in urban landscape scene analysis,aiming to provide efficient and high-precision images for autonomous vehicles.Cityscape Scene Parsing Algorithms.The main research content of this paper is as follows:(1)Two key difficulties in urban landscape scene parsing are deeply studied,namely,the preservation of high resolution and the acquisition of high-level semantic information.In-depth study of the design of convolutional neural network,the feature map near the input end has high resolution and high semantic information,but lacks edge and detail information;the feature map near the output end has a low resolution,loses semantic information,but acquires a lot It is difficult to meet the needs of scene analysis.Therefore,this paper proposes a convolutional neural network that combines hollow convolution and pyramid pooling to solve these two difficulties.(2)For the problem of multi-scale object segmentation,an improved high-resolution network combined with empty space pyramid pooling is proposed.First,half of the residual module of the high-resolution network is improved by depth-separable convolution,and the feature extraction network is based on this;secondly,the hole-space pyramid pooling module with improved hole rate stacking is designed to obtain multiple scales The receptive field of the model enhances the ability of the model to extract multi-scale object features;finally,a dualchannel parallel decoder structure is designed to aggregate all the advantages and obtain highquality segmentation images.Finally,with 34.8M parameters on the City Scapes dataset,82.1%MIo U was obtained,and the forward prediction speed reached 10 FPS.(3)Based on the research on the improved high-resolution network combined with hollow space pyramid pooling,a high-resolution scene parsing network is proposed for the context information aggregation problem.Firstly,the high-resolution network is improved by the hole-separable convolution superimposed with the three-level hole rate,and the feature extraction network is based on this;secondly,a multi-stage fusion upsampling structure is designed to make full use of the improved high-resolution The output feature maps of the four parallel paths of the network;finally,an improved pyramid pooling module that can adapt to input images of different sizes is designed to aggregate global context information and local context information of different sizes,optimize the segmentation results,and obtain highquality Split graph.Finally,with 16.4M parameters on the City Scapes dataset,83.3% MIo U was obtained,and the forward prediction speed reached 14 FPS.(4)Apply the above algorithm to the actual application,design and realize a city landscape scene analysis system based on deep learning for automatic driving,and carry out system design and system implementation completely according to the specifications of software engineering.This paper expounds in detail the whole content of system requirement analysis,system design,system realization and system test.The scene analysis system includes modules such as user management,training model,scene analysis,data set management,and algorithm management.It meets the needs of users for scene analysis of different data sets.